Automated hybrid breeding

Abstract

A modified plant suitable for automated hybrid breeding, a preparation method therefor, and a corresponding automatic hybrid breeding method.

Description

Automated hybrid breeding Technical Field

[0001] The present invention relates to the field of biotechnology, and more particularly to the field of plant breeding. The present invention provides a modified plant suitable for automated hybrid breeding, a preparation method thereof, and a corresponding automated hybrid breeding method.

[0002] Background of the Invention

[0003] For a century, the discovery of heterosis has revolutionized crop breeding. Heterosis is a ubiquitous phenomenon in crop breeding, whereby the first generation (F1) hybrids exhibit superior yield, stress tolerance, quality, adaptability, or other traits compared to their parents. First discovered in maize (Zea mays), heterosis has been used to increase yields in a variety of major crops, including corn, rice (Oryza sativa), soybean (Glycine max), and tomato (Solanum lycopersicum), by 50-100%, 55%, 47%, and 20-50%, respectively, contributing to global food security. Hybrid breeding has made it possible to improve crop growth characteristics, enhance resistance to biotic and abiotic stresses, and increase yield. The global F1 hybrid seed market was valued at US$22.9 billion in 2021 and is projected to reach US$41.88 billion by 2030, with a compound annual growth rate (CAGR) of 6.9% during the forecast period. Furthermore, the global tomato seed market was valued at US$1.22 billion in 2022 alone and is projected to reach US$2.15 billion by 2030. However, modern hybrid breeding still faces numerous challenges. For example, crop hybridization efficiency is low, which is closely related to the diversity of crop propagation systems, making manual pollination time-consuming and labor-intensive.

[0004] The success of hybrid breeding in a particular crop depends primarily on the characteristics of its reproductive system. Crops can be categorized by their reproductive system, including cross-pollinated, actively cross-pollinated, and self-pollinated crops. Cross-pollinated crops are often self-incompatible, meaning that certain individuals cannot be fertilized by their own pollen. In contrast, self-pollinated crops can be fertilized by their own pollen and produce self-fertilized offspring. Actively cross-pollinated crops can undergo both self- and cross-pollination. While wild plants evolve through random mutation and natural selection, domesticated plants have evolved through intense artificial selection, which has enabled the selection and preservation of favored traits. Once favorable traits emerge and are selected, humans gradually establish inbred lines to stabilize these traits. This genetic stability is often accompanied by changes in crop reproductive systems, most notably changes in floral organ morphology. For example, the relative position of the stigma and stamens may shift, from an exposed stigma adapted for outcrossing to a retracted stigma more suitable for self-pollination. This morphological change in the retracted stigma helps preserve the "pure" inbred line traits for continuous cultivation, but it hinders the outcrossing requirements of modern breeding. For commercial hybrid seed production, the stamens of the flowers need to be manually removed to expose the stigma. Because stamen removal must be performed before pollen matures, this process may damage other floral organs, thereby reducing the success rate of cross-pollination and fertilization. In addition, due to the long time it takes for this trait to be introgressed into different genetic backgrounds, the applicability of the male sterile lines that have been discovered is limited. Therefore, there is an urgent need to create male sterile lines with exposed stigmas in a variety of excellent genetic backgrounds for scalable hybrid breeding.

[0005] Tomato is one of the most valuable specialty crops. As a self-pollinating crop, commercial tomato seeds are mostly hybrids. However, the highest cost of hybrid tomato seed production is labor expenditure, as the stigma is completely enclosed within the stamen group, making hybrid seed production entirely dependent on manual detasseling and pollination. Over 50 naturally occurring or induced male sterile tomato mutants have been discovered in various genetic backgrounds. These plants can be categorized as either "structural male sterility" or "functional male sterility" based on the cause of male sterility. The former often manifests as defects in stamen development, while the latter is typically caused by anther indehiscence or defects in pollen development. The stamenless (sl) mutant series, consisting of five mutants of uncertain allelic relationship (sl-1 to sl-5), exhibits typical structural male sterility, characterized by the transformation of stamens into carpel-like structures. The stigma of these mutants is exposed due to malformed and twisted stamens. The carpelization phenotype of the sl mutants resembles the transformation of stamens into carpels caused by deletion of class B genes in the ABC model of floral development. The ABC model is based on studies in Arabidopsis thaliana and Antirrhinum majus, in which petal and stamen fates are determined by APETALA3 (AP3) and PISTILLATA (PI) in Arabidopsis or by the antirrhinum orthologs DEFICIENS (DEF) and GLOBOSA (GLO) in antirrhinum. Tomato has four class B MADS-box genes: AP3 (Tomato AP3 [TAP3], also known as DEF) and Tomato MADS box gene 6 (TM6, also known as TDR6) in the AP3 clade, and GLOBOSA (GLO1, also known as PI and TPIB) and PISTILLATA (Tomato PI [TPI], also known as GLO2) in the PI clade. Mutations in DEF result in the stamen-defective phenotypes of sl and tap3 mutants.

[0006] Previous studies have reported two male-sterile tomato mutants with exserted stigmas, sl2 and 7B-1. Map-based cloning mapped the candidate gene for 7B-1 to chromosome 6, where it is linked to GLO2, a homolog of the Arabidopsis B-class MADS-box gene PI, which specifies stamen identity. Although recombination between 7B-1 and GLO2 has occurred, the complex structure of GLO2's second intron and 3' UTR has made it difficult to amplify its full-length sequence and confirm its identity as the 7B-1 candidate gene. In summary, these B-class tomato gene mutants have the potential for hybrid breeding. However, they also face challenges such as the time-consuming, multi-generational nature of traditional backcrossing to change genetic backgrounds or growth defects, which severely limit their application in tomato hybrid breeding.

[0007] An ideal male sterile tomato line requires precisely modified floral morphology to prevent self-pollination while also expressing the favorable agronomic traits of the desired genetic background. Two strategies exist to achieve this goal: one is to restore cultivated tomatoes to their ancestral stigma length while simultaneously rendering them pollen sterile, but this may require altering genes with complex chromosomal structures. The other is to genetically engineer B-class genes of the ABC model to achieve male sterility resulting in exserted stigmas due to stamen defects. In traditional exserted stigma sterile lines, mutations in ABC model genes result in significant phenotypic changes, but are difficult to implement in commercial production environments due to low adaptability or additional yield defects.

[0008] Artificial intelligence (AI) is increasingly being used in agriculture, primarily in field cultivation, water and fertilizer management, weed control and pesticide spraying, disease monitoring, and harvesting. Most of these applications involve image acquisition, big data analysis, and automated operations for objects of varying sizes. In recent years, pollination robots have been developed for tomato and kiwifruit (Actinidia chinensis). By gently vibrating the flowers, the pollination robots disperse pollen, enabling self-pollination and fertilization. In contrast, current robots are not suitable for hybrid breeding of seeds with encapsulated stamens, as this requires precise manipulation to simultaneously open the stamen cone and complete pollination without damaging the stigma. The stigma encapsulated within the stamen group presents challenges for robotic detasseling and increases production costs.

[0009] Summary of the Invention

[0010] The combination of biotechnology and artificial intelligence (BT+AI) has enormous potential. Biotechnology can produce AI-friendly plant morphologies, thereby creating a better plant-machine "interface" and enabling intelligent, precise, and scalable automated breeding operations. The present invention uses gene editing to enable plants with enclosed stigmas, such as tomatoes and soybeans, to acquire an exposed stigma phenotype. This phenotype is particularly suitable for automated pollination, which is beneficial for the application of the plant in hybrid breeding. The present invention further improves the artificial intelligence algorithm and robotic arm control method of the automated pollination robot, achieving efficient automated pollination of plants with an exposed stigma phenotype.

[0011] In one aspect, the present invention provides a method for producing a modified plant, comprising targeted modification of at least one class B gene of an endogenous floral development ABC model of the plant, wherein the modification results in a stigma exsertion phenotype in the modified plant.

[0012] In another aspect, the present invention provides a modified plant, wherein at least one class B gene of the endogenous floral development ABC model of the plant is modified, wherein the modification results in an exserted stigma phenotype in the modified plant. For example, the modified plant is produced by the method of the present invention.

[0013] In another aspect, the present invention provides use of the modified plants of the present invention in cross-breeding, such as automated cross-breeding.

[0014] In another aspect, the present invention provides a method for processing an image containing a flower and a method and system for automated pollination. These methods and systems are particularly applicable to automated pollination of the modified plants of the present invention.

[0015] In another aspect, the present invention provides a method for automated plant hybrid breeding using an automatic pollination robot, the method comprising:

[0016] i) providing a plant having a flower with an exserted stigma phenotype as a hybrid mother plant;

[0017] ii) enabling the automatic pollination robot to approach the plant and obtain an image of the plant;

[0018] iii) enabling an automatic pollination robot to process the image using an artificial intelligence method, detect the flowers of the plant, and determine the three-dimensional position of the stigma; and

[0019] iv) causing an automatic pollination robot to pollinate the stigmas in the flowers of the hybrid female plant with pollen from the hybrid male plant.

[0020] BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1. Rapid creation of structurally male sterile lines in tomato by editing the GLO2 gene. (A) The retracted stigma of cultivated tomato (Solanum lycopersicum) and the exserted stigma of wild tomato (Solanum pimpinellifolium) are dominant in modern hybrid breeding and used for hybrid seed production. White arrows indicate the location of the stigma. Scale bar, 2 mm. (B) Schematic diagram of the target site targeting the GLO2 coding region (top) and the resulting edited mutant form (bottom). Black squares and narrower black squares indicate exons and UTRs, respectively. Red font indicates sgRNA, black boxes indicate PAM sequences, and cyan font indicates mutations. (C) Schematic diagram of the target site targeting the GLO2 noncoding region (second intron and 3' UTR) (top) and the resulting edited mutant form (bottom). Blue arrows indicate primers. (D) Schematic diagram of the target site targeting the second intron region of GLO2 (top) and the resulting edited mutant form (bottom). Two repeat regions within the second intron are indicated by yellow squares, and sequences associated with the miniature inverted repeat transposable element (MITE) are indicated by light blue squares. (E) Images show the inflorescence (left), stamen cone structure (center), and pollen fertility (right). Arrows indicate exposed stigmas. Scale bars: inflorescence (10 mm), floral organs (1 mm), and pollen stain (100 μm). (F) Statistical data on flowering time in Ailsa Craig and glo2 mutants. (G, H) Statistical data on plant height (from the base of the stem to the third inflorescence) (G) and number of flowers on the first inflorescence (H) in Ailsa Craig and glo2 mutants. (I, J) The ratio of pistil to stamen length in different glo2 mutants indicates the degree of stigma exposure. Statistical analysis was performed using a two-tailed, two-sample t-test. Sample size and p-value are indicated on the statistical graphs.

[0022] Figure 2. Fruit and seed phenotypes of different glo2 mutants. (A) Representative images of a single fruit (top) and all fruits of the first three ears (bottom). Fruits of male sterile lines (glo2-dele, glo2-inver, and glo2-in2-a2) were obtained by hybridization using AC as the male parent. The deformed fruit rates of glo2-dele, glo2-inver, and glo2-in2-a2 were 2.17%, 1.98%, and 13.95%, respectively. The five-pointed star indicates the location of the deformed fruit. Scale bars, top (1 cm), bottom (5 cm). (BD) Fruit set rate of the first two ears per plant (B), fruit weight of the first three ears (C), and total fruit yield (D). When the fruit content of each ear exceeded 95% at the color breaking and maturity stages, the first three ears were harvested for yield quantification. (E) Representative images of stamen defect phenotypes of varying degrees: normal stamens (NO), stamens with carpelized structures on the front (CS), stamens with carpelized structures and externally attached ovules (EO), and stamens completely converted to carpels (TC). (F) Quantitative data showing the degree of stamen weakening, as well as carpels and fruit ventricles, for various glo2 mutant alleles. Sample size is indicated alongside the statistical percentages. (G) Germination rate test (top) and appearance comparison (bottom) of hybrid seeds harvested from three male sterile lines (glo2-dele, glo2-inver, and glo2-in2-a2) with AC seeds. Scale bar, 1 cm. (H, J) Statistical data showing seed number per fruit (H), germination rate (I), and 100-seed weight (J) of hybrid fruits harvested from three male sterile lines (glo2-dele, glo2-inver, and glo2-in2-a2) after crossing with AC. Unless otherwise indicated, mutants in blue are hybrid fruits or hybrid seeds. Statistical analysis was performed using a two-tailed two-sample t-test. Sample sizes and p-values ​​are indicated on the statistical graphs.

[0023] Figure 3. AI-driven autonomous pollination of male sterile tomatoes with exposed stigmas. (A, B) Comparison of the time cost (A) and fruit set rate (B) of hybridization of Alisa Craig and glo2-inver by artificial pollination. The hybridization was performed independently by three skilled personnel, with 10 hybridization events in each group. The fruit set rate refers to the fruit set rate of flowers on a single inflorescence after hybridization. (C) Schematic diagram of tomato flower and stigma recognition. The image was resized to W×H×3 before being input into YOLACT_Orient. In this work, the pixel resolution ratio of W and H was set to 550. The output parameters (x, y, w, h, c, m, o) represent the confidence that the current box is a tomato flower (c), the box range of the tomato flower (x, y, w, h), m prototype masks for instance segmentation, and the direction of the stigma (o). (D) Comparison of the flower detection accuracy and time cost of different methods. (E) The improved "YOLACT_Orient" method and Comparison of flower orientation classification using the Bayesian Classifier (NBC). P: precision, R: recall. (F) Pseudo-binocular ranging strategy. Using a pseudo-binocular ranging strategy, Speeded-Up Robust Features (SURF) and Randomized Sample Consensus (RANSAC) are combined to calculate the three-dimensional position of the pistil in the "eye-hand camera" coordinate system. (G) Recognition errors in the X, Y, and Z dimensions, as well as the time cost of stigma position identification. A two-tailed, two-sample t-test was performed. Sample size and p-value are indicated on the graph.

[0024] Figure 4. Cruise robot for tomato pollination. (AC) Schematic diagram (A) and photos (B, C) show the structure and hardware of the pollination robot. The pollination robot consists of seven components: (1) Carrier / ShiHe-MR1000 base, (2) Ultra-wide-band (UWB) positioning module, (3) pollination arm, (4) pollination gripper, (5) Camera / RealSense-435i, (6) pollination brush, and (7) pollen container (see Methods for details). (D) Arrangement of plants in the greenhouse and the robot's cruise route. The light green rectangle represents the plant range, the dark green circle represents the planting location, the dark blue solid line represents the vehicle's cruise route, and the blue rectangle with numbered corners is the Ultra-wide-band (UWB) gateway. (E) Schematic diagram of the vehicle's cruise and operation on a plant during the robot pollination process, where W, L, and H in the left figure represent the width, length, and height range of the plant, respectively, and X, Y, and Z in the right figure represent the working range of the pollination robot arm. (F) Comparison of the robot's pollination range and the range of the plant's flowers. The data in the middle bar graph show the ideal working range of the pollination robot arm in the Z, Y, and X axes, while the data in the right bar graph show the actual working range of the pollination robot arm (ideal values ​​minus a 10 cm operating error). (G) Circular spiral servo trajectory on the two-dimensional flower plane, sequentially passing through 36 key points (12 points on three circles) (left), and the servo trajectory in three-dimensional space. The red solid arrow indicates the direction of the robot arm's movement (right). (H) Workflow of the pollination servo (K = 36). (I) Distribution of successful pollination servo times obtained from a total of 320 flower-stigma contact experiments. (J) Comparison of fruit set rate between manual and robotic pollination. Five autonomous pollination experiments were conducted on glo2-inver plants using the robot, using 5-18 flowers per group, and the fruit set rate was quantified for each experiment. Manual pollination of wild-type plants served as a control. (K) Photographs showing the step-by-step process of the robotic pollination. (L) Comparison of time costs between manual and robotic pollination. Manual pollination of wild-type plants served as a control. A two-tailed, two-sample t-test was performed. Sample size and p-value are indicated on the graph.

[0025] Figure 5. Combining GEAIR with de novo domestication allows for rapid breeding of superior lines in robotically pollinated progeny. (A) Schematic diagram of the GEAIR-driven rapid breeding process for stress-tolerant tomatoes. (B, C) Greenhouse phenotype (B) and representative individual plants (C) of the best inbred lines from the F4 progeny of a cross between glo2-inver (Ailsa Craig background) and S. pimpinellifolium. Scale bar, 1 cm. (D, G) Statistics showing total fruit yield per plant (D), fruit weight (E), Brix (F), and lycopene content (G) of selected inbred lines. Data are mean ± SD, analyzed using a two-tailed, two-sample t-test. Sample size and p-value are indicated on the graph. (HP) Content of flavor volatiles in fruits of selected inbred lines. Samples were prepared from six red-ripe fruits randomly collected from three plants, and three technical replicates were used for each sample for statistical analysis. Data are mean ± SD, analyzed using a two-tailed, two-sample t-test. p-value is indicated on the graph.

[0026] Figure 6. Some Solanaceae relatives are candidates for male sterility creation through editing of class B MADS-box genes. (A) Phylogenetic tree of representative cultivated Solanaceae species and some wild relatives, including Petunia hybrid (petunia), Nicotiana benthamiana (tobacco), Capsicum annuum (pepper), Physalis pruinosa (mushroom), Solanum melongena (eggplant), Solanum tuberosum (potato), Solanum pennellii (wild tomato), Solanum chilense (wild tomato), Solanum habrochaites (wild tomato), Solanum pimpinellifolium (wild tomato), Solanum galapagense (wild tomato), and Solanum lycopersicum (cultivated tomato). Data from TimeTree (http: / / www.timetree.org). Also included are the corresponding geological age (bottom) and flower morphology (right) for each species. PLE, Pleistocene; RUP, Ruppelian; AQT, Aquitanian; LAN, Langeian; SER, Serravallian; MES, Messinian; ZAN, Zankelian; PIA, Piacenzian; GEL, Grassian; LOW, Early Pleistocene; MID, Middle Pleistocene; MYA, million years ago. (B) Comparison of stamen and stigma morphology in representative Solanaceae species, including fused stamens (orange boxes), intermediate morphology (yellow boxes), and separate stamens (blue boxes). (C) Phylogenetic analysis of GLO2 homologs in 24 representative Solanaceae species with different stamen types. St, stamen; Ca, carpel.

[0027] Figure 7. Design of the GEAIR breeding system and phenotypes of various glo2 mutants, Supplementary Figure 1. (A) Schematic diagram of changes in floral organs and pollination types during tomato domestication and breeding. (B) Schematic diagram of the GEAIR design. Se, sepal; Pe, petal; St, stamen; Ca, carpel. (C) PCR gel electrophoresis showing multiple deletion and inversion alleles in glo2 mutants. Primers spanning the entire target region were used to amplify the target fragments, see Figures 1B and 1C. (D) RT-qPCR analysis showing GLO2 transcript levels in flower buds (0.5 cm to 1 cm in length) of different mutant alleles. GLO2 expression levels were referenced to tomato Ubiquitin gene expression. Data are mean ± SD. (E) Representative images of flower, stigma, and stamen morphology after petal removal. (F) RT-qPCR showing the relative expression levels of three other class B MADS-box genes in various glo2 mutants. Expression levels were referenced to tomato Ubiquitin gene expression. Data are presented as mean ± standard deviation and two-tailed two-sample t-test was performed. P values ​​are indicated in the bar graphs.

[0028] Figure 8. Male sterility and morphological changes in fruit and floral organs of the glo2 mutant under different growth conditions, supplementary to Figures 1 and 2. (A) Phenotypes of glo2 mutants grown in the field. Images were taken 80 days after transplanting. White circles indicate inflorescences. Scale bar, 10 cm. (B) Fruits of glo2 mutants grown in the greenhouse. Images were taken 80 days after transplanting. Brackets and numbers indicate the number of leaves within the sympodial growth cycle. Red arrows indicate the sympodial inflorescence and the first inflorescence. Scale bar, 10 cm. (C) Close-up images of different stamen defects. White arrows indicate carpel structures. Stamens with carpel structures on the front (CS), stamens with carpel structures and externally attached ovules (EO), and stamens completely converted to carpels (TC). Scale bars, 1 mm (center TC), 200 μm (lower TC, CS, and EO). (D) Ventricle of mature glo2 mutant fruits. Scale bar, 5 cm. (E) Cross-sections of carpels from various glo2 mutants. White arrows indicate carpels. Scale bar, 200 μm.

[0029] Figure 9. Stigma orientation, detection, calculation, and robotic pollination process interface, supplementary to Figures 3 and 4. (A) Representative images of different stigma orientations. The stigma points to the left, right, front, top, and bottom. Green lines indicate the five orientation ranges, red arrows indicate the marked directions, and the blue circle indicates the center of the flower. (B) Images in the columns labeled Original Image, Ground Truth, Mask-RCNN, DETR, and YOLACT_Orient were captured by a RealSense D435i camera, annotated by the user, and used by Mask-RCNN, DETR, and YOLACT_Orient to detect, segment, and determine flower orientation. Dark red rectangles indicate detected flower regions, and bright red arrows indicate flower orientation. Flower segmentation regions derived by Mask-RCNN, DETR, and YOLACT_Orient are represented by masks of different colors. The capital letters L, R, F, U, and D in the red boxes represent left, right, front, top, and bottom, respectively. The numbers in the red boxes represent the confidence level of the flower orientation. In this study, only flowers with exposed stigmas were considered pollination targets. (C) Detailed representation of the human view after pseudo-binocular odometry and motion planning. (D) The interface in (C) shows the X, Y, and Z coordinates of the robot end-effector in the operating coordinate system. These values ​​are highlighted in red rectangles. (E) Detailed representation of the human view when the end-effector contacts the stigma after processing the circular spiral servo trajectory. (F) The interface in (E), where the values ​​in the red rectangle record the X, Y, and Z values ​​of the end-effector in the robot's operating coordinate system. The difference in the X, Y, and Z values ​​within the red rectangles in (D) and (F) demonstrates the level of accuracy achieved in position identification using pseudo-binocular odometry technology. (G) Schematic diagram of the robotic pollination process.

[0030] Figure 10. Production of F1 hybrid tomatoes using the GEAIR system and phenotypic analysis, supplementary to Figure 4. (A) Schematic diagram of GEAIR-enhanced F1 hybrid tomato production. (BL) Representative images of parental and F1 plants, including the whole plant (left), single ear of fruit (top right), individual fruit (middle right), and all fruit per plant (first six ears, bottom right). Scale bars: 10 cm (whole plant, all fruit per plant), 1 cm (ear, individual fruit). (MO) Statistics of total fruit yield per plant (first six ears), fruit weight, and Brix ratio for parental and F1 plants. Two-tailed, two-sample t-tests were used. Sample size and p-values ​​are indicated on the graph. ** p<0.01, *** p<0.001. (P) Heterosis analysis for all five hybrid combinations. Intermediate and superior heterosis were calculated, and red numbers indicate positive heterosis.

[0031] Figure 11. Phenotypes of parental lines, de novo domesticated plants, and GEAIR-mediated hybrid offspring, supplement to Figure 5. (AD) Representative images show the phenotypes of parental plants and offspring of the glo2-inver (AC) mutant hybridized with the de novo domesticated wild tomato S. pimpinellifolium. Three different growth habits (plant structures) were isolated from the F2 offspring, namely sp (limited growth), sp5g (day-neutral and early flowering), and sp sp5g (limited growth, day-neutral and early flowering). Scale bars: 10 cm (A), 1 cm (BD). (E, F) Statistics show the flowering time and fruit yield per inflorescence of the F2 generation of three different plant structures. A two-tailed two-sample t-test was performed. The sample size and p-value are indicated on the figure, ns, not significant; ** p<0.01; *** p<0.001. (G) Representative leaves of powdery mildew-infected plants of the parental lines and the F2 population. White circles indicate the location of infection, and the percentage of infected plants of each genotype is shown in the figure. The scale bar is 1 cm. (H, I) (H) shows all fruits of representative plants of AC (left) and F3 offspring (right) grown in saline-alkali soil (salinity: 0.3%, pH: 8.0) in the greenhouse. Statistics (I) show the number of fruits per plant. The picture was taken 3 months after transplanting. The scale bar is 1 cm. The data were subjected to a two-tailed two-sample t-test. The sample size and p-value are indicated on the figure. * p<0.05, *** p<0.001.

[0032] Figure 12 Protein sequence analysis of class B genes from representative Solanaceae plants, supplementary to Figure 6. (AD) Sequence analysis of GLO2 (A), GLO1 (B), DEF (C), and TM6 (D) homologous proteins from representative Solanaceae plants with different stamen types. GLO2 is divided into two groups: group A represents stamen fusion and group B represents stamen separation. Intermediate morphological types (yellow vertical lines) belong to one of these two groups, suggesting the involvement of other class B MADS-box proteins. The differentiation of the other three class B MADS-box proteins, GLO1, DEF, and TM6, was then analyzed. The results showed that, with the exception of GLO1, the protein sequence analysis of DEF and TM6 showed similarities to GLO2, reflecting the differentiation of stamen types. Group A is indicated by red asterisks, vertical lines, and underlines, while group B is indicated by blue asterisks, vertical lines, and underlines.

[0033] Figure 13. Creation of a structurally male sterile line by editing a class B MADS-box gene in soybean. (A) External morphology of a soybean (Glycine max) flower, showing the sepals and petals. The petals consist of five petals, including a flagellum, two wing petals, and two connate keels. The stamens and pistils are tightly enclosed within the keels. Scale bar, 1 mm. (B) Schematic diagram of the creation of a structurally male sterile line by editing a class B MADS-box gene in soybean using CRISPR technology. (C) Schematic diagram of sgRNAs targeting the coding region of a class B MADS-box gene (top) and mutant allele information (bottom). Black squares and narrower black squares represent exons and UTRs, respectively. Red font indicates the sgRNA, and black-bordered boxes indicate the PAM sequence. Cyan font indicates mutations. (D-I) Representative photographs of flowers (column 1), flowers with calyx removed (column 2), calyxes (column 3), petals (column 4), stamens (column 5), and pistils (column 6) of wild-type soybean, gmpi1 gmpi3+ / -, gmpi3, gmpi1 gmpi3, gmap3b gmtm6b, and gmap3a gmap3b gmtm6b mutants. Scale bar, 1 mm. (J) Leaves of wild-type soybean and the gmap3a gmap3b gmtm6b mutant. Scale bar, 1 cm. (K) Seed setting of wild-type soybean and the gmap3a gmap3b gmtm6b mutant under natural conditions. Scale bar, 1 cm. (L) Seed setting of wild-type soybean and the gmap3a gmap3b gmtm6b mutant after artificial cross-pollination. Scale bar, 1 cm.

[0034] FIG14 is a schematic diagram showing that the three-dimensional space in which the pollination device is driven to search for the stigma is a cylinder.

[0035] FIG15 is a schematic block diagram showing a system for automated pollination according to the present invention.

[0036] FIG16 . Schematic perspective view of a system for automated pollination in the form of a pollination robot.

[0037] Detailed Description of the Invention

[0038] 1. Definition

[0039] In the present invention, unless otherwise indicated, the scientific and technical terms used herein have the meanings commonly understood by those skilled in the art. In addition, the terms and laboratory procedures related to protein and nucleic acid chemistry, molecular biology, cell and tissue culture, microbiology, and immunology used herein are terms and routine procedures widely used in the corresponding fields. For example, the standard recombinant DNA and molecular cloning techniques used in the present invention are well known to those skilled in the art and are more fully described in the following literature: Sambrook, J., Fritsch, EF and Maniatis, T., Molecular Cloning: A Laboratory Manual; Cold Spring Harbor Laboratory Press: Cold Spring Harbor, 1989 (hereinafter referred to as "Sambrook"). At the same time, in order to better understand the present invention, definitions and explanations of relevant terms are provided below.

[0040] As used herein, the term "and / or" encompasses all combinations of items connected by the term, and should be treated as if each combination had been individually listed herein. For example, "A and / or B" encompasses "A," "A and B," and "B." For example, "A, B, and / or C" encompasses "A," "B," "C," "A and B," "A and C," "B and C," and "A and B and C."

[0041] When the word "comprising" is used herein to describe a sequence of a protein or nucleic acid, the protein or nucleic acid may be composed of the sequence, or may have additional amino acids or nucleotides at one or both ends of the protein or nucleic acid, but still have the activity described in the present invention. In addition, it is clear to those skilled in the art that the methionine encoded by the start codon at the N-terminus of the polypeptide may be retained in certain practical situations (for example, when expressed in a specific expression system), but it does not substantially affect the function of the polypeptide. Therefore, when describing a specific polypeptide amino acid sequence in the specification and claims of this application, although it may not contain a methionine encoded by a start codon at the N-terminus, a sequence containing the methionine is also covered, and accordingly, its encoding nucleotide sequence may also contain a start codon; and vice versa.

[0042] "Polynucleotide," "nucleic acid sequence," "nucleotide sequence," or "nucleic acid fragment" are used interchangeably and are single-stranded or double-stranded polymers of RNA or DNA that optionally contain synthetic, non-natural, or altered nucleotide bases. Nucleotides are referred to by their single-letter designations as follows: "A" for adenosine or deoxyadenosine (RNA or DNA, respectively), "C" for cytidine or deoxycytidine, "G" for guanosine or deoxyguanosine, "U" for uridine, "T" for deoxythymidine, "R" for purine (A or G), "Y" for pyrimidine (C or T), "K" for G or T, "H" for A or C or T, "I" for inosine, and "N" for any nucleotide. Although nucleotide sequences herein may be presented as DNA sequences (including T), when reference is made to RNA, one skilled in the art can readily determine the corresponding RNA sequence (i.e., replacing T with U).

[0043] "Polypeptide," "peptide," and "protein" are used interchangeably herein to refer to a polymer of amino acid residues. The terms apply to amino acid polymers in which one or more amino acid residues is an artificial chemical analog of a corresponding naturally occurring amino acid, as well as to naturally occurring amino acid polymers. The terms "polypeptide," "peptide," "amino acid sequence," and "protein" may also include modified forms including, but not limited to, glycosylation, lipid attachment, sulfation, gamma-carboxylation of glutamic acid residues, hydroxylation, and ADP-ribosylation.

[0044] Sequence "identity" has a meaning recognized in the art, and the percentage of sequence identity between two nucleic acid or polypeptide molecules or regions can be calculated using published techniques. Sequence identity can be measured along the entire length of a polynucleotide or polypeptide or along a region of the molecule. (See, for example: Computational Molecular Biology, Lesk, AM, ed., Oxford University Press, New York, 1988; Biocomputing: Informatics and Genome Projects, Smith, DW, ed., Academic Press, New York, 1993; Computer Analysis of Sequence Data, Part I, Griffin, AM, and Griffin, HG, eds., Humana Press, New Jersey, 1994; Sequence Analysis in Molecular Biology, von Heinje, G., Academic Press, 1987; and Sequence Analysis Primer, Gribskov, M. and Devereux, J., eds., M Stockton Press, New York, 1991). While there are many methods to measure the identity between two polynucleotides or polypeptides, the term "identity" is well known to those of skill in the art (Carrillo, H. & Lipman, D., SIAM J Applied Math 48: 1073 (1988)).

[0045] In peptides or proteins, suitable conservative amino acid substitutions are known to those skilled in the art and can generally be made without altering the biological activity of the resulting molecule. Generally, those skilled in the art recognize that single amino acid substitutions in non-essential regions of a polypeptide do not substantially alter biological activity (see, e.g., Watson et al., Molecular Biology of the Gene, 4th Edition, 1987, The Benjamin / Cummings Pub.co., p. 224).

[0046] As used herein, an "expression construct" refers to a vector, such as a recombinant vector, suitable for expressing a nucleotide sequence of interest in an organism. "Expression" refers to the production of a functional product. For example, expression of a nucleotide sequence can refer to the transcription of the nucleotide sequence (e.g., transcription to produce mRNA or functional RNA) and / or translation of RNA into a precursor or mature protein.

[0047] The "expression construct" of the present invention can be a linear nucleic acid fragment, a circular plasmid, a viral vector, or, in some embodiments, can be an RNA (such as mRNA) that can be translated.

[0048] An "expression construct" of the present invention may comprise regulatory sequences and a nucleotide sequence of interest from different sources, or regulatory sequences and a nucleotide sequence of interest from the same source but arranged in a manner different from that normally found in nature.

[0049] "Regulatory sequence" and "regulatory element" are used interchangeably to refer to nucleotide sequences located upstream (5' non-coding sequences), within, or downstream (3' non-coding sequences) of a coding sequence and that influence the transcription, RNA processing or stability, or translation of the associated coding sequence. Regulatory sequences may include, but are not limited to, promoters, translation leader sequences, introns, and polyadenylation recognition sequences.

[0050] "Promoter" refers to a nucleic acid fragment that is capable of controlling the transcription of another nucleic acid fragment. In some embodiments of the present invention, a promoter is a promoter that is capable of controlling the transcription of a gene in a cell, whether or not it is derived from the cell. A promoter can be a constitutive promoter, a tissue-specific promoter, a developmentally regulated promoter, or an inducible promoter. A "constitutive promoter" refers to a promoter that generally causes a gene to be expressed in most cell types under most circumstances. "Tissue-specific promoter" and "tissue-preferred promoter" are used interchangeably and refer to a promoter that is primarily, but not necessarily exclusively, expressed in one tissue or organ, and may also be expressed in one specific cell or cell type. A "developmentally regulated promoter" refers to a promoter whose activity is determined by developmental events. An "inducible promoter" selectively expresses an operably linked DNA sequence in response to endogenous or exogenous stimuli (environmental, hormone, chemical signal, etc.). Examples of promoters include, but are not limited to, polymerase (pol) I, pol II, or pol III promoters. When used in plants, the promoter can be cauliflower mosaic virus 35S promoter, maize Ubi-1 promoter, wheat U6 promoter, rice U3 promoter, maize U3 promoter, rice actin promoter.

[0051] As used herein, the term "operably linked" refers to the connection of a regulatory element (e.g., but not limited to, a promoter sequence, a transcription termination sequence, etc.) to a nucleic acid sequence (e.g., a coding sequence or an open reading frame) such that transcription of the nucleotide sequence is controlled and regulated by the transcriptional regulatory element. Techniques for operably linking regulatory element regions to nucleic acid molecules are known in the art.

[0052] "Introducing" a nucleic acid molecule (e.g., a plasmid, a linear nucleic acid fragment, RNA, etc.) or a protein into an organism refers to transforming the cells of the organism with the nucleic acid or protein so that the nucleic acid or protein can function in the cell. "Transformation" as used in the present invention includes stable transformation and transient transformation. "Stable transformation" refers to the introduction of an exogenous nucleotide sequence into the genome, resulting in the stable inheritance of the exogenous gene. Once stably transformed, the exogenous nucleic acid sequence is stably integrated into the genome of the organism and any successive generations thereof. "Transient transformation" refers to the introduction of a nucleic acid molecule or protein into a cell to perform a function without the stable inheritance of the exogenous gene. In transient transformation, the exogenous nucleic acid sequence is not integrated into the genome.

[0053] As used herein, the term "plant" includes whole plants and any offspring, cells, tissues, or parts of plants. The term "plant part" includes any part of a plant, including, for example, but not limited to: seeds (including mature seeds, immature embryos without seed coats, and immature seeds); plant cuttings; plant cells; plant cell cultures; plant organs (e.g., pollen, embryos, flowers, fruits, buds, leaves, roots, stems, and related explants). Plant tissues or plant organs can be seed, callus, or any other plant cell colony organized into structural or functional units. Plant cells or tissue cultures can regenerate plants with the physiological and morphological characteristics of the plant from which the cells or tissues were derived, and can regenerate plants with substantially the same genotype as the plant. In contrast, some plant cells cannot regenerate to produce plants. The regenerable cell in plant cell or tissue culture can be an embryo, protoplast, meristematic cell, callus tissue, pollen, leaf, anther, root, root tip, silk, flower, kernel, ear, cob, husk, or stem.

[0054] Plant "progeny" include any subsequent generation of the plant.

[0055] "Trait" refers to the physiological, morphological, biochemical or physical characteristics of a cell or organism. "Agronomic traits" specifically refer to measurable indicators of crop plants, including but not limited to: leaf greenness, seed yield, growth rate, total biomass or accumulation rate, fresh weight at maturity, dry weight at maturity, fruit yield, seed yield, plant total nitrogen content, fruit nitrogen content, seed nitrogen content, plant vegetative tissue nitrogen content, plant total free amino acid content, fruit free amino acid content, seed free amino acid content, plant vegetative tissue free amino acid content, plant total protein content, fruit protein content, seed protein content, plant vegetative tissue protein content, herbicide resistance and drought resistance, nitrogen uptake, root lodging, harvest index, stem lodging, plant height, ear height, ear length, disease resistance, cold resistance, salt resistance and tiller number, etc.

[0056] 2. Methods for Producing Modified Plants

[0057] In one aspect, the present invention provides a method for producing a modified plant, the method comprising modifying at least one endogenous class B gene of the ABC model of floral development of the plant, the modification resulting in a stigma exsertion phenotype in the modified plant.

[0058] As used herein, "exposed stigma" refers to a plant flower in which the stigma is not obscured or enclosed by other floral organs, such as petals and / or stamens. "Exposed stigma" includes a stigma that is at least partially exposed, preferably fully exposed. Exposed stigmas can be easily identified and located by machines for automated pollination. The modified plants described herein are particularly suitable for automated crossbreeding due to their exposed stigma phenotype.

[0059] The class B genes are genes that determine petal and stamen development in plants, and may be homologous genes of Arabidopsis thaliana APETALA3 (AP3) or PISTILLATA (PI); or homologous genes of snapdragon DEFICIENS (DEF) and GLOBOSA (GLO). Based on genomic analysis and sequence comparison, those skilled in the art can easily identify candidate class B genes in plants.

[0060] In some embodiments, the plant is a plant whose stigma is naturally covered or wrapped by other floral organs such as petals and / or stamens, that is, the stigma in the flowers of wild-type plants that have not been modified is covered or wrapped by other floral organs such as petals and / or stamens, preferably completely covered or wrapped.

[0061] In some embodiments, the plant is a Solanaceae plant. In some embodiments, the plant is selected from Solanum lycopersicum (tomato), Nicotiana benthamiana (tobacco), Capsicum annuum (pepper), Physalis pruinosa (mushroom), Solanum melongena (eggplant), Solanum tuberosum (potato), Solanum pennellii, Solanum chilense, Solanum habrochaites, Solanum pimpinellifolium, Solanum galapagense, and Petunia hybrid (petunia). Preferably, the plant is Solanum lycopersicum (tomato).

[0062] In some embodiments, the plant can be from a different cultivar of Solanum lycopersicum (tomato). In some embodiments, the plant is tomato cultivar Ailsa Craig, M82, Beijing 1, TS545, TS181, TS590. Preferably, the plant is tomato cultivar Ailsa Craig.

[0063] In some embodiments, the at least one class B gene is GLO2, such as a tomato GLO2 gene. In some embodiments, the GLO2 gene encodes a GLO2 protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence of SEQ ID NO: 1. In some embodiments, the GLO2 gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 2. In some embodiments, the GLO2 gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 3.

[0064] In some embodiments, the plant is a legume (Fabaceae). In some embodiments, the plant is selected from soybean (Glycine max), peanut, broad bean, pea, red bean, mung bean, cowpea, kidney bean and lentil. In some preferred embodiments, the plant is soybean (Glycine max).

[0065] In some embodiments, the at least one Class B gene is selected from the group consisting of soybean GmPI1, GmPI2, GmPI3, GmPI4, GmAP3a, GmAP3b, GmTM6a, and GmTM6b genes. In some preferred embodiments, the at least one Class B gene is selected from the group consisting of soybean GmAP3a, GmAP3b, and GmTM6b genes. In some preferred embodiments, the at least one Class B gene includes soybean GmAP3a, GmAP3b, and GmTM6b genes.

[0066] In some embodiments, the GmPI1 gene encodes a GmPI1 protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence of SEQ ID NO: 19. In some embodiments, the GmPI1 gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 20. In some embodiments, the GmPI1 gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 21.

[0067] In some embodiments, the GmPI3 gene encodes a GmPI3 protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence of SEQ ID NO: 24. In some embodiments, the GmPI3 gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 25. In some embodiments, the GmPI3 gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 26.

[0068] In some embodiments, the GmAP3a gene encodes a GmAP3a protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence of SEQ ID NO: 29. In some embodiments, the GmAP3a gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 30. In some embodiments, the GmAP3a gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 31.

[0069] In some embodiments, the GmAP3b gene encodes a GmAP3b protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence of SEQ ID NO: 33. In some embodiments, the GmAP3b gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 34. In some embodiments, the GmAP3b gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 35.

[0070] In some embodiments, the GmTM6a gene encodes a GmTM6a protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence of SEQ ID NO: 37. In some embodiments, the GmTM6a gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 38. In some embodiments, the GmTM6a gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 39.

[0071] In some embodiments, the GmTM6b gene encodes a GmTM6b protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence of SEQ ID NO: 41. In some embodiments, the GmTM6b gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 42. In some embodiments, the GmTM6b gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 43.

[0072] In some embodiments, the modification is a targeted modification. In some embodiments, the modification is a substitution, deletion and / or addition of one or more nucleotides. In some embodiments, the modification is a rearrangement of a fragment of the gene, such as an inversion.

[0073] In some embodiments, the modification is performed by introducing into the plant a gene editing system that targets the at least one class B gene.

[0074] The gene editing system that can be used in the present invention can be various gene editing systems known in the art, as long as it can perform targeted genome editing in plants. The gene editing system can be a gene editing system based on CRISPR, ZFN or TALEN. Preferably, the gene editing system is a gene editing system based on CRISPR.

[0075] The CRISPR gene editing system can include a CRISPR nuclease and at least one guide RNA, and / or an expression construct encoding the nuclease and guide RNA. The CRISPR nuclease and guide RNA can form a complex that targets and / or cleaves a genomic target sequence based on complementarity between the guide RNA and the genomic target sequence.

[0076] "CRISPR nuclease" can be derived from a Cas9 nuclease, including a Cas9 nuclease or a functional variant thereof. The Cas9 nuclease can be a Cas9 nuclease from a different species, such as spCas9 from Streptococcus pyogenes (S. pyogenes) or SaCas9 derived from Staphylococcus aureus (S. aureus). "Cas9 nuclease" and "Cas9" are used interchangeably herein and refer to an RNA-guided nuclease comprising a Cas9 protein or a fragment thereof (e.g., a protein comprising an active DNA cleavage domain of Cas9 and / or a gRNA binding domain of Cas9). Cas9 is a component of the CRISPR / Cas (clustered regularly interspaced short palindromic repeats and related systems) genome editing system that can target and cut a DNA target sequence to form a DNA double-strand break (DSB) under the guidance of a guide RNA.

[0077] "CRISPR nuclease" can also be derived from Cpf1 nuclease, including Cpf1 nuclease or functional variants thereof. The Cpf1 nuclease can be Cpf1 nuclease from different species, such as Cpf1 nuclease from Francisella novicida U112, Acidaminococcus sp. BV3L6, and Lachnospiraceae bacterium ND2006.

[0078] Useful "CRISPR nucleases" can also be derived from Cas3, Cas8a, Cas5, Cas8b, Cas8c, Cas10d, Cse1, Cse2, Csy1, Csy2, Csy3, GSU0054, Cas10, Csm2, Cmr5, Cas10, Csx11, Csx10, Csf1, Csn2, Cas4, C2c1, C2c3 or C2c2 nucleases, for example, including these nucleases or functional variants thereof.

[0079] As used herein, “guide RNA” and “gRNA” are used interchangeably and refer to an RNA molecule that can form a complex with a CRISPR effector protein and can target the complex to a target sequence due to a certain homology with the target sequence. The guide RNA targets the target sequence by base pairing with the complementary strands of the target sequence. For example, the gRNA used by the Cas9 nuclease or its functional variants is generally composed of crRNA and tracrRNA molecules that are partially complementary to form a complex, wherein the crRNA comprises a guide sequence (also known as a seed sequence) that has sufficient homology to the target sequence so as to hybridize with the complementary strand of the target sequence and guide the CRISPR complex (Cas9+crRNA+tracrRNA) to specifically bind to the target sequence sequence. However, it is known in the art that single guide RNA (sgRNA) can be designed, which contains the features of crRNA and tracrRNA at the same time. The gRNA used by the Cpf1 nuclease or its functional variants is generally composed of only mature crRNA molecules, which may also be referred to as sgRNA. It is within the capabilities of those skilled in the art to design a suitable gRNA based on the CRISPR nuclease used and the target sequence to be edited.

[0080] The gene editing system described in the present invention may also encompass so-called base editing systems or prime editing systems.

[0081] Methods that can be used to introduce the gene editing system of the present invention into plants include, but are not limited to, gene gun method, PEG-mediated protoplast transformation, Agrobacterium-mediated transformation, plant virus-mediated transformation, pollen tube channel method, and ovary injection method.

[0082] In the method of the present invention, target sequence modification can be achieved simply by introducing or generating a gene editing system into plant cells, and the modification is stably inherited, without the need for stably transforming the gene editing system into plants. This avoids potential off-target effects of a stable gene editing system and prevents integration of exogenous nucleotide sequences into the plant genome, thereby enhancing biosafety.

[0083] In some embodiments, the introduction includes converting the gene editing system to isolated plant cells or tissues, and then regenerating the converted plant cells or tissues into complete plants. In other embodiments, the gene editing system can be converted to specific parts on the complete plant, such as leaves, stem tips, pollen tubes, young ears or hypocotyls. This is particularly suitable for the transformation of plants that are difficult to regenerate through tissue culture.

[0084] The gene editing system can target any region of the Class B gene, as long as it can cause a decrease in the expression / function of the Class B gene. For example, the gene editing system can target the endogenous coding sequence of the Class B gene, resulting in mutation or incomplete translation of the protein. Alternatively, the gene editing system can target the non-coding region of the Class B gene, resulting in reduced expression or incomplete translation of the protein.

[0085] In some embodiments, the modification results in the plant expressing a truncated protein encoded by a class B gene. In other embodiments, the modification results in downregulation of expression of a class B gene in the plant by at least 2-fold, at least 5-fold, at least 10-fold, at least 20-fold, at least 30-fold, at least 40-fold, at least 50-fold, at least 100-fold, or more.

[0086] In some embodiments, the plant is a Solanaceae plant such as a tomato. In some embodiments, the gene editing system targets the non-coding region of the GLO2 gene. In some specific embodiments, the gene editing system targets the second intron and / or 3'UTR (3' untranslated region) of the GLO2 gene. In some specific embodiments, the gene editing system targets the second intron of the GLO2 gene. In some specific embodiments, the gene editing system targets the second intron and 3'UTR of the GLO2 gene.

[0087] In some embodiments, the gene editing system comprises a target sequence selected from one of SEQ ID NOs: 4-11 or a combination thereof.

[0088] In some embodiments, the modification results in a mutated GLO2 gene of one of SEQ ID NOs: 12-18.

[0089] In some embodiments, the modification results in the plant expressing a truncated GLO2 protein. In other embodiments, the modification results in downregulation of the expression of the GLO2 gene in the plant by at least 2-fold, at least 5-fold, at least 10-fold, at least 20-fold, at least 30-fold, at least 40-fold, at least 50-fold, at least 100-fold, or more.

[0090] In some embodiments, the plant is a leguminous plant such as soybean. In some embodiments, the gene editing system targets the coding region of the at least one class B gene (GmPI1, GmPI2, GmPI3, GmPI4, GmAP3a, GmAP3b, GmTM6a, GmTM6b or any combination thereof), preferably the 5' end of the coding sequence.

[0091] In some embodiments, the gene editing system comprises a target sequence selected from one of SEQ ID NOs: 22-23, 27-28, 32, 36, 40, and 44, or a combination thereof.

[0092] In some embodiments, the modification results in a mutated class B gene of one of SEQ ID NOs: 45-56, or any combination thereof.

[0093] In some embodiments, the modification results in mutated GmAP3a, GmAP3b, and GmTM6b genes. In some embodiments, the modification results in a mutated GmAP3a gene as set forth in SEQ ID NO: 49 or 50, a mutated GmAP3b gene as set forth in SEQ ID NO: 51 or 52, and a mutated GmTM6b gene as set forth in SEQ ID NO: 55 or 56.

[0094] In some embodiments, the modification does not affect pistil development in the modified plant. In some embodiments, the modification results in impaired stamen development in the modified plant.

[0095] In some embodiments, the modification results in a male sterile phenotype in the modified plant. In some embodiments, the modified plant has a complete male sterile phenotype. In some embodiments, the modified plant does not produce fertile pollen.

[0096] In some embodiments, the modification is homozygous or heterozygous, preferably homozygous.

[0097] In some embodiments, the modified plant has comparable flowering time, plant height, and / or inflorescence architecture as compared to an unmodified wild-type plant.

[0098] In some embodiments, the modified plant has comparable fruit quality as the unmodified wild-type plant after artificial pollination (e.g., with pollen from a corresponding unmodified wild-type plant). For example, the modified plant has a proportion of deformed fruit of less than about 10%, less than about 5%, less than about 4%, less than about 3%, less than about 2%, less than about 1%, or even less.

[0099] In some embodiments, the modified plant has comparable seed number, 100-seed weight, and / or seed germination rate as the unmodified wild-type plant following artificial pollination (eg, with pollen from a corresponding unmodified wild-type plant).

[0100] In some embodiments, the modified plant has comparable fruit yield as the unmodified wild-type plant following artificial pollination (eg, pollen from a corresponding unmodified wild-type plant).

[0101] 3. Modified plants

[0102] In another aspect, the present invention provides a modified plant, wherein at least one class B gene of the endogenous floral development ABC model of the plant is modified, and the modification results in a stigma exsertion phenotype in the modified plant.

[0103] The class B genes are genes that determine petal and stamen development in plants, and may be homologous genes of Arabidopsis thaliana APETALA3 (AP3) or PISTILLATA (PI); or homologous genes of snapdragon DEFICIENS (DEF) and GLOBOSA (GLO). Based on genomic analysis and sequence comparison, those skilled in the art can easily identify candidate class B genes in plants.

[0104] In some embodiments, the plant is a plant whose stigma is naturally covered or wrapped by other floral organs such as petals and / or stamens, that is, the stigma in the flowers of wild-type plants without the modification is covered or wrapped by other floral organs such as petals and / or stamens.

[0105] In some embodiments, the plant is a Solanaceae plant. In some embodiments, the plant is selected from Solanum lycopersicum (tomato), Nicotiana benthamiana (tobacco), Capsicum annuum (pepper), Physalis pruinosa (mushroom), Solanum melongena (eggplant), Solanum tuberosum (potato), Solanum pennellii, Solanum chilense, Solanum habrochaites, Solanum pimpinellifolium, Solanum galapagense, and Petunia hybrid (petunia). Preferably, the plant is Solanum lycopersicum (tomato).

[0106] In some embodiments, the plant can be from a different cultivar of Solanum lycopersicum (tomato). In some embodiments, the plant is tomato cultivar Ailsa Craig, M82, Beijing 1, TS545, TS181, TS590. Preferably, the plant is tomato cultivar Ailsa Craig.

[0107] In some embodiments, the at least one class B gene is GLO2, such as a tomato GLO2 gene. In some embodiments, the GLO2 gene encodes a GLO2 protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence of SEQ ID NO: 1. In some embodiments, the GLO2 gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 2. In some embodiments, the GLO2 gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 3.

[0108] In some embodiments, the plant is a legume (Fabaceae). In some embodiments, the plant is selected from soybean (Glycine max), peanut, broad bean, pea, red bean, mung bean, cowpea, kidney bean and lentil. In some preferred embodiments, the plant is soybean (Glycine max).

[0109] In some embodiments, the at least one Class B gene is selected from the group consisting of soybean GmPI1, GmPI2, GmPI3, GmPI4, GmAP3a, GmAP3b, GmTM6a, and GmTM6b genes. In some preferred embodiments, the at least one Class B gene is selected from the group consisting of soybean GmAP3a, GmAP3b, and GmTM6b genes. In some preferred embodiments, the at least one Class B gene includes soybean GmAP3a, GmAP3b, and GmTM6b genes.

[0110] In some embodiments, the GmPI1 gene encodes a GmPI1 protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence of SEQ ID NO: 19. In some embodiments, the GmPI1 gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 20. In some embodiments, the GmPI1 gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 21.

[0111] In some embodiments, the GmPI3 gene encodes a GmPI3 protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence of SEQ ID NO: 24. In some embodiments, the GmPI3 gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 25. In some embodiments, the GmPI3 gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 26.

[0112] In some embodiments, the GmAP3a gene encodes a GmAP3a protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence of SEQ ID NO: 29. In some embodiments, the GmAP3a gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 30. In some embodiments, the GmAP3a gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 31.

[0113] In some embodiments, the GmAP3b gene encodes a GmAP3b protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence of SEQ ID NO: 33. In some embodiments, the GmAP3b gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 34. In some embodiments, the GmAP3b gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 35.

[0114] In some embodiments, the GmTM6a gene encodes a GmTM6a protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence of SEQ ID NO: 37. In some embodiments, the GmTM6a gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 38. In some embodiments, the GmTM6a gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 39.

[0115] In some embodiments, the GmTM6b gene encodes a GmTM6b protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence of SEQ ID NO: 41. In some embodiments, the GmTM6b gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 42. In some embodiments, the GmTM6b gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 43.

[0116] In some embodiments, the modification is a substitution, deletion and / or addition of one or more nucleotides. In some embodiments, the modification is a rearrangement of a fragment of the gene such as an inversion.

[0117] The modification can be located in any region of the Class B gene, as long as it can result in reduced expression / function of the Class B gene. For example, the modification can be located in the endogenous coding sequence of the Class B gene, resulting in mutation or incomplete translation of the protein. Alternatively, the modification can be located in the non-coding region of the Class B gene, resulting in reduced expression or incomplete translation of the protein.

[0118] In some embodiments, the modification results in the plant expressing a truncated protein encoded by a class B gene. In other embodiments, the modification results in downregulation of expression of a class B gene in the plant by at least 2-fold, at least 5-fold, at least 10-fold, at least 20-fold, at least 30-fold, at least 40-fold, at least 50-fold, at least 100-fold, or more.

[0119] In some embodiments, the plant is a Solanaceae plant such as a tomato. In some embodiments, the modification may be located in a non-coding region of the GLO2 gene. In some specific embodiments, the modification may be located in the second intron and / or 3'UTR (3' untranslated region) of the GLO2 gene. In some specific embodiments, the modification may be located in the second intron of the GLO2 gene. In some specific embodiments, the modification may be located between the second intron and 3'UTR of the GLO2 gene.

[0120] In some embodiments, the modification results in a mutated GLO2 gene of one of SEQ ID NOs: 12-18.

[0121] In some embodiments, the modification results in the plant expressing a truncated GLO2 protein. In some embodiments, the modification results in downregulation of expression of the GLO2 gene in the plant by at least 2-fold, at least 5-fold, at least 10-fold, at least 20-fold, at least 30-fold, at least 40-fold, at least 50-fold, at least 100-fold, or more.

[0122] In some embodiments, the plant is a leguminous plant such as soybean. In some embodiments, the modification is located in the coding region of the at least one class B gene (GmPI1, GmPI2, GmPI3, GmPI4, GmAP3a, GmAP3b, GmTM6a, GmTM6b or any combination thereof), preferably the 5' end of the coding sequence.

[0123] In some embodiments, the modification results in a mutated class B gene of one of SEQ ID NOs: 45-56, or any combination thereof.

[0124] In some embodiments, the modification results in mutated GmAP3a, GmAP3b, and GmTM6b genes. In some embodiments, the modification results in a mutated GmAP3a gene as set forth in SEQ ID NO: 49 or 50, a mutated GmAP3b gene as set forth in SEQ ID NO: 51 or 52, and a mutated GmTM6b gene as set forth in SEQ ID NO: 55 or 56.

[0125] In some embodiments, the modification is homozygous or heterozygous, preferably homozygous.

[0126] In some embodiments, the modified plant has normal pistil development. In some embodiments, the modified plant has impaired stamen development.

[0127] In some embodiments, the modified plant has a male sterile phenotype. In some embodiments, the modified plant has a completely male sterile phenotype. In some embodiments, the modified plant does not produce fertile pollen.

[0128] In some embodiments, the modified plant has comparable flowering time, plant height, and / or inflorescence architecture as compared to an unmodified wild-type plant.

[0129] In some embodiments, the modified plant has comparable fruit quality as the unmodified wild-type plant after artificial pollination (e.g., with pollen from a corresponding unmodified wild-type plant). For example, the modified plant has a proportion of deformed fruit of less than about 10%, less than about 5%, less than about 4%, less than about 3%, less than about 2%, less than about 1%, or even less.

[0130] In some embodiments, the modified plant has comparable seed number, 100-seed weight, and / or seed germination rate as the unmodified wild-type plant following artificial pollination (eg, with pollen from a corresponding unmodified wild-type plant).

[0131] In some embodiments, the modified plant has comparable fruit yield as the unmodified wild-type plant following artificial pollination (eg, pollen from a corresponding unmodified wild-type plant).

[0132] In some embodiments, the modified plant is produced by the method of the present invention described above.

[0133] In another aspect, the present invention provides use of the modified plants of the present invention in hybrid breeding. In some embodiments, the modified plants of the present invention are used in automated hybrid breeding.

[0134] 4. Methods and Systems for Automated Pollination

[0135] To complete automated pollination, an automatic pollination robot needs to accurately detect and position the flower's stigma. It then moves a soft pollination device coated with pollen near the stigma, gently operating the device to complete pollination. Currently, pollination robots face problems such as low flower stigma positioning accuracy and low pollination success rates, as well as potential damage to crops (particularly the stigma) during the pollination process. Therefore, a method and system for precisely positioning pollination stigmas and flexibly controlling pollination are needed. This method and system are particularly suitable for automated pollination of the modified plants of the present invention.

[0136] In one aspect, the present invention provides a method for automated pollination, comprising:

[0137] detecting the presence of stigmas of crops to be pollinated;

[0138] Acquire images of the crop stigma at two different positions, and obtain a three-dimensional location of the crop stigma based on matching feature points between the images;

[0139] Based on the three-dimensional positioning of the crop stigma, a pollination device carrying pollen is driven to move in front of the crop stigma; and

[0140] The pollination device is driven to move sequentially between a plurality of points in a set three-dimensional space including the above-mentioned three-dimensional positioning until the pollination device coats the pollen onto the crop stigma.

[0141] In another aspect, the present invention also provides a system for automated pollination, comprising:

[0142] Removable base;

[0143] a robotic arm assembly mounted on the base;

[0144] a pollination device, a camera, and a pollen box each detachably attached to the robotic arm assembly; and

[0145] A controller configured to:

[0146] controlling a camera to acquire images of the crop to detect the presence of a stigma of the crop to be pollinated;

[0147] Controlling a camera to acquire images of the crop stigma at two different positions, and obtaining a three-dimensional location of the crop stigma based on matching feature points between the images;

[0148] The driving base moves based on the three-dimensional positioning of the crop stigma, so that the pollination device and the pollen box carried by the robotic arm assembly move to the front of the crop stigma; and

[0149] The pollination device is driven to move sequentially between a plurality of points in a set three-dimensional space including the above-mentioned three-dimensional positioning until the pollination device applies the pollen obtained from the pollen box to the crop stigma.

[0150] According to the method for automated pollination of the present invention, the first step is to detect the presence of stigmas on the crop to be pollinated. This can be done by using a camera (such as the camera included in the automated pollination system described later) to photograph the crop and analyzing the captured image to detect the presence of stigmas. It should be noted that automated pollination targets crops with exposed stigmas. Therefore, if stigmas are present in the crop image, methods such as deep learning can be used to detect their presence. Methods for detecting the presence of crop stigmas are well known in the art and are not the focus of the present invention, so they will not be discussed in detail here.

[0151] After detecting the presence of the crop stigma to be pollinated, the next step performed by the method of the present invention is to obtain the stigma's three-dimensional position in the camera coordinate system. Specifically, the camera can capture images of the stigma at two different positions, and the stigma's three-dimensional position in the camera coordinate system can be determined based on matching feature points between the images.

[0152] Using binocular vision for 3D positioning is well known in the art. This approach typically involves simultaneously capturing images of a target using two cameras at different locations. The 3D position of the target relative to the cameras is determined by matching feature points between the two images. However, in the field of automated crop pollination, deploying two cameras directly at the end of the system increases calibration costs and the likelihood of the cameras and other accessories scratching or even damaging crops, particularly stigmas, during movement.

[0153] To overcome the above shortcomings, the present invention adopts a "pseudo-binocular" three-dimensional positioning method. Specifically, the present invention equips the system for automated pollination with only one camera. Through precise manipulation at the end of the robotic arm, the camera moves to obtain images of the stigma at two different positions, and obtains the three-dimensional precise positioning of the stigma based on the matching of feature points between the images. Therefore, compared with a binocular system using two cameras, the "pseudo-binocular" three-dimensional positioning method of the present invention achieves at least equivalent positioning accuracy, while also reducing the calibration cost and space occupied by the system, and reducing the possibility of scratching or even damaging crops, especially the stigma.

[0154] The camera may photograph the column head at one position and then move horizontally to another nearby position to photograph the column head. The distance between the two positions is, for example, 10 cm, but the present invention is not limited thereto.

[0155] Image feature point matching algorithms are well known in the field of image processing. For example, feature point matching algorithms such as SIFT (Scale-invariant feature transform) and SURF (Speeded Up Robust Features) can be combined with the RANSAC (Random Sample Consensus) algorithm to obtain pre-matched feature points between pseudo-binocular image pairs. However, the present invention is not limited to this, and other feature point matching algorithms can also be used to obtain the three-dimensional location of the column head. It should be noted that the three-dimensional location obtained in this case is the location (three-dimensional coordinates) of the column head in the camera coordinate system. Specifically, the three-dimensional location can be the three-dimensional coordinates of the center of the column head in the camera coordinate system.

[0156] When matching feature points, other parts of the crop besides the petals, as well as objects in the crop's surroundings, may constitute background noise and adversely affect the accuracy of the match, thereby reducing the three-dimensional positioning accuracy of the stigma. To this end, an image of the petal region, including the stigma, can be cropped from the image captured by the camera, and feature point matching can be performed only based on the cropped image, thereby removing background noise and improving the three-dimensional positioning accuracy of the stigma. Furthermore, since the size of the cropped image is reduced, the speed of calculating the three-dimensional positioning of the stigma is also accelerated. According to the applicant's experiments, the average time required to calculate the three-dimensional positioning of the stigma is approximately 0.045 seconds. Thus, the stigma can be quickly positioned in three dimensions.

[0157] The next step performed by the method of the present invention is: based on the three-dimensional positioning of the stigma, the system drives the pollination device carrying pollen to move to the front of the stigma. At this time, it is necessary to convert the three-dimensional positioning of the stigma in the camera coordinate system into its three-dimensional positioning in the system coordinate system of the system for automated pollination. As long as the camera is calibrated after being installed in the system, the specific position of the camera in the system can be known, or the three-dimensional positioning of the camera relative to the pollination device of the system (such as a pollination brush) can be known. Therefore, as long as the three-dimensional positioning of the stigma relative to the camera is obtained, it is equivalent to obtaining the three-dimensional positioning of the stigma relative to the pollination device. In order to complete the subsequent pollination operation, it is necessary to first move the pollination device along the stigma toward the corresponding direction to the vicinity of the stigma to be pollinated. For example, based on the three-dimensional positioning of the stigma relative to the pollination device, the pollination device is driven to move to the front near the stigma, for example, 10 cm directly in front of the stigma, but the present invention is not limited to this.

[0158] In order to move the pollination device to the vicinity of the stigma to be pollinated, the positioning of the pollination device itself in the environment should also be clear. Taking the pollination of multiple rows of crops (such as tomato flowers) in a greenhouse as an example, wireless gateways can be installed off the ground at the four corners of the greenhouse, and the system carrying the pollination device can be connected to the network and can know its position and direction in the greenhouse in real time, and the position of the root position of each crop in each row in the greenhouse is also known. Therefore, as long as the presence of a stigma to be pollinated on a crop is detected and the three-dimensional positioning of the stigma is obtained by matching feature points, a path can be planned for it based on the current positioning of the networked system to move to the front near the stigma. Any base that can be used to move the pollination device of the present invention to a specified pollination target, such as an automatic roaming base, can be used for the purpose of the present invention.

[0159] The positioning algorithm based on feature point matching may have certain positioning errors. Taking the experiment of the applicant to perform three-dimensional positioning on the exposed stigmas of 128 randomly selected tomato flowers as an example, by comparing the three-dimensional positioning of the stigma obtained by the "pseudo-binocular" three-dimensional positioning method with the three-dimensional positioning of the stigma actually measured, it was found that the stigma had an average positioning error of 0.50cm, 0.85cm and 0.53cm in the three axes of the three-dimensional coordinate system X, Y, and Z, respectively, and a maximum positioning error of 1.33cm, 2.22cm and 1.29cm in the three axes, respectively. Due to the existence of positioning errors, if the pollination device is directly driven to move to the three-dimensional positioning of the stigma obtained by the positioning algorithm based on feature point matching, the pollination device may not contact the stigma, resulting in unsuccessful pollination.

[0160] In order to increase the success rate of pollination, the method of the present invention proposes that after the pollination device moves to the front near the stigma, the pollination device carrying pollen is driven to move sequentially between multiple points in a set three-dimensional space, including the three-dimensional positioning of the stigma obtained according to the above steps, until the pollination device coats the pollen onto the stigma.

[0161] That is, the method according to the present invention adopts a "coarse to fine" strategy to achieve pollination, first moving the pollination device to the vicinity of the stigma, and then moving it through multiple points to achieve a fine search of the stigma.

[0162] The aforementioned three-dimensional space can be understood as the spatial range for the pollination device to retrieve the stigma, which can be set as a regular three-dimensional space, such as a sphere, a cuboid, a cube, a cylinder, etc., or it can also be set as an irregular three-dimensional space. Preferably, the three-dimensional space is set as a three-dimensional space that covers the maximum positioning error of the three-dimensional positioning of the stigma obtained by matching the feature points in the three-dimensional coordinate system X, Y, and Z. Taking the maximum positioning errors of 1.33 cm, 2.22 cm, and 1.29 cm in the three-dimensional coordinate system X, Y, and Z obtained by the applicant's experiment as an example, the above-mentioned three-dimensional space can be set as a three-dimensional space formed by taking the three-dimensional positioning of the stigma obtained by matching the feature points as the center, extending 1.33 cm in the positive and negative directions of the X axis, 2.22 cm in the positive and negative directions of the Y axis, and 1.29 cm in the positive and negative directions of the Z axis.

[0163] Therefore, as long as the pollination device is driven to move sequentially between enough points in the aforementioned three-dimensional space, it can be ensured that when the pollination device moves to a certain point, the pollen will be successfully applied to the stigma to complete pollination. The multiple points in the aforementioned three-dimensional space are, for example, selected to be evenly distributed in the three-dimensional space, and the number of points can be determined based on the error size of the three-dimensional positioning of the stigma obtained by matching the feature points, that is, the number of points is adjustable to improve the pollination efficiency while ensuring the success rate of pollination. At the same time, once successful pollination is determined, the pollination device is stopped from being driven to continue moving to other points in the three-dimensional space. Instead, the existence of other crop stigmas to be pollinated is re-detected, and the pollination device is driven to move to the vicinity of the newly detected crop stigma and a new pollination process is started. Thereby, the pollination efficiency is further improved.

[0164] Because the stigma is more likely to be located at or near the 3D location of the stigma obtained through feature point matching, and less likely to be located at the outer contour of the 3D space, the pollination device can be driven to first move to the stigma's 3D location obtained through feature point matching, and then gradually move to the outer contour of the 3D space through multiple points from near to far. Once successful pollination is confirmed, the pollination device is stopped and continues to move to other points in the 3D space. This further improves pollination efficiency.

[0165] The applicant has discovered that if the aforementioned three-dimensional space is set as an irregular solid space, selecting the point within the space to which the pollination device is to move becomes relatively complex, and it is also not conducive to path planning for the pollination device to move sequentially between multiple points. Therefore, it is preferable to set the three-dimensional space as a regular solid space.

[0166] Specifically, the present invention has discovered that constructing the aforementioned three-dimensional space as a cylinder is particularly advantageous. This cylinder encompasses the maximum positioning error of the stigma's three-dimensional location, obtained through feature point matching, in the three-dimensional coordinate system X, Y, and Z. Figure 14 shows a schematic diagram of a cylinder in which the three-dimensional space used to drive the pollination device to search for the stigma. The center of mass C of the cylinder is the stigma's three-dimensional location, obtained through feature point matching.

[0167] In order to facilitate the retrieval of the stigma, the pollination device is driven to move sequentially along a plurality of points on the annular spiral line in at least two circular cross sections of the cylinder until the stigma is pollinated.

[0168] Specifically, referring to FIG14 , the two circular surfaces S1 and S2 of the cylinder are located in a vertical plane (parallel to the Z-axis direction). In order to make the pollination device contact the stigma to complete pollination, the pollination device can first be driven to move to the first circular surface S1 of the cylinder. On this surface, the pollination device is servo-moved along the annular spiral line, that is, it moves sequentially between multiple points on the annular spiral line in the vertical plane. If the pollination device is not detected to be in contact with the stigma on surface S1, the pollination device is driven to move a certain distance in the direction from the center of circular surface S1 to the center of circular surface S2, for example, to the circular surface (shown as surface S3 in the figure) that forms the cross section of the cylinder and is located between surfaces S1 and S2. On this surface, The pollination device continues to be servo-moved along the circular spiral line, that is, it moves in sequence between multiple points on the circular spiral line in the vertical plane. If the pollination device is still not detected to be in contact with the stigma in the cross-section, the pollination device is continued to be driven to move to another circular surface (another cross-section of the cylinder formed between surfaces S1 and S2, or surface S2), in which the pollination device continues to be servo-moved along the circular spiral line, that is, it moves in sequence between multiple points on the circular spiral line in the vertical plane, until successful pollination is determined, and then the driving of the pollination device for the above-mentioned servo movement along the circular spiral line is stopped.

[0169] That is, the pollination device moves sequentially along multiple points along at least two parallel circular spirals in the three-dimensional space of the cylinder until successful pollination is determined. In this manner, the pollination device only needs to translate to the position of each circular spiral and then move sequentially along multiple points along a circular spiral in a vertical plane at that position to ensure that the pollination device successfully applies pollen to the stigma during the process.

[0170] Figure 14 shows three spirals constructed in the three-dimensional space of a cylinder, but the present invention is not limited to this. Depending on the maximum positioning error of the stigma's three-dimensional positioning obtained through feature point matching, an appropriate number of circular spirals can be selected within the cylinder. Specifically, the span between adjacent circular spirals can be appropriately selected based on this maximum positioning error. The number of points on each circular spiral can also be selected accordingly, and the number and distribution of points on each circular spiral can be consistent. This simplifies path planning for the pollination device while ensuring successful application of pollen to the stigma.

[0171] In order to determine whether the pollination device has successfully applied pollen to the stigma, the method adopted is to use a camera to obtain an image containing the stigma and the pollination device every time the pollination device moves to a point, and determine whether the two are in contact based on their relative positions in the image. It should be noted that since this application only uses one camera, there may be a possibility of misjudgment when judging whether pollination is successful based on only one image obtained by it. This misjudgment may be caused by factors such as interference from other objects around the stigma. Therefore, in order to improve the accuracy of judging whether pollination is successful, the method of the present invention proposes to determine that pollination is successful only when the pollination device and the stigma are judged to be in contact twice in a row.

[0172] To expedite the process of determining whether the pollination device has successfully applied pollen to the stigma, the present invention utilizes a lightweight network architecture based on Inception-v3. Each time an image containing the stigma and the pollination device is acquired, it can quickly confirm whether the stigma and the pollination device are in contact. The Inception-v3 network architecture is well known in the art and will not be described in detail here. (Of course, any visual-tactile processing method for detecting whether the stigma is touched by the pollination device is applicable to the present invention.)

[0173] The present invention also provides a system for automated pollination. FIG15 shows a schematic block diagram of a system 100 for automated pollination according to the present invention, the system 100 comprising:

[0174] a movable base 110;

[0175] A robotic arm assembly 120 mounted on the base 110;

[0176] a pollination device 130 , a camera 140 , and a pollen box 150 detachably attached to the robotic arm assembly 120 ; and

[0177] A controller 160 configured to:

[0178] controlling the camera 140 to acquire images of the crop to detect the presence of stigmas of the crop to be pollinated;

[0179] Controlling the camera 140 to acquire images of the crop stigma at two different positions, and obtaining a three-dimensional location of the crop stigma based on matching of feature points between the images;

[0180] The driving base 110 moves based on the three-dimensional positioning of the crop stigma, so that the pollination device 130 and the pollen box 150 carried by the robot arm assembly 120 move to the front of the crop stigma; and

[0181] The pollination device 130 is driven to move sequentially between a plurality of points in a set three-dimensional space including the aforementioned three-dimensional positioning, until the pollination device 130 applies the pollen obtained from the pollen box 150 to the crop stigma.

[0182] That is, the controller 160 is configured to control the hardware components of the system 100 to execute the method according to the present invention. Figure 15 shows that the controller 160 is in communication with each hardware component in the dashed box to control the operation of each hardware component.

[0183] The system 100 of the present invention can be implemented in the form of an automatic pollination robot. Figure 16 shows a schematic perspective view of the system 100 for automated pollination in the form of an automatic pollination robot, wherein the base 110, the robotic arm assembly 120, the pollination device 130, the camera 140 and the pollen box 150 of the system 100 are shown. Among them, the base 110, the robotic arm assembly 120 and the camera 140 can adopt corresponding commercially available products, the pollination device 130 (not shown in Figure 16) can adopt, for example, a wool pen or a wool brush, which can be grasped by the end of the robotic arm assembly 120, such as a mechanical claw 121, so as to be detachably attached to the robotic arm assembly 120, and the pollination device 130 can be moved to the pollen box 150 (not shown in Figure 16) to obtain pollen and can be moved to the stigma of the crop to be pollinated by the robotic arm assembly 120.

[0184] The present invention also provides a machine-readable storage medium storing executable instructions. When the instructions are executed by a processor module, the method for automated pollination according to the present invention is implemented.

[0185] Applicants implemented the method and system of the present invention on tomato flowers with exposed stigmas and conducted extensive experiments. They found that using a "pseudo-binocular" three-dimensional positioning method and a "coarse-to-fine" pollination strategy, they achieved efficient and highly successful automated pollination. However, the method and system of the present invention are not limited to tomato flowers; they are also applicable to the pollination of various other crops with exposed stigmas.

[0186] 5. Processing Methods for Images Containing Flowers

[0187] The biggest challenge in automated pollination lies in detecting and locating the stigma of the target flower, as it can be obscured by leaves, branches, fruit, and other flowers. Stigmas can also be oriented in different directions, and the exact angle of orientation determines the approach trajectory that the robotic pollen brush must follow to achieve pollination. Manually designing such flower-stigma estimation algorithms based on specialized knowledge is extremely difficult. Therefore, there remains a need for effective image-based methods for detecting and locating stigmas in flowers that can be applied to automated pollination.

[0188] In one aspect, the present invention provides a method for processing an image containing a flower, comprising:

[0189] A) Input image containing flowers;

[0190] B) Send the input image to the backbone network for feature extraction to obtain a feature map;

[0191] C) sending the obtained feature map to the detection branch to obtain prediction coefficients, and sending the obtained feature map to the segmentation branch to obtain a prototype mask, wherein the obtained prediction coefficients include coefficients representing the flower orientation; and

[0192] D) The obtained prototype mask is combined with the prediction coefficients, and the output result is obtained together with the coefficient representing the flower orientation through the cropping and threshold modules.

[0193] In some embodiments, in step A), the image is pre-processed, wherein the size of the input image is adjusted to W×H×3. In some embodiments, the pixel resolution of W and H is set to 550. However, the pixel resolution setting of the present invention is not limited thereto.

[0194] Based on the YOLACT network framework (Bolya, D., et al., (2022). YOLACT++Better Real-Time Instance Segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(2), 1108-1121. DOI: 10.1109 / TPAMI.2020.3014297), the present invention adds a flower orientation classification network, which is called YOLACT_Orient in the present invention. YOLACT_Orient can not only simultaneously detect and segment the range of flowers, but also infer the orientation of flowers.

[0195] The feature main network marked as c1-c5 in Figure 3C is built by generating masks using Fully Convolutional Networks (FCN).

[0196] Higher resolution pyramid prototypes are achieved by using Feature Pyramid Networks (FPN) to connect c3 and p3 with the same size. FPN then continuously upsamples to a quarter of the input image size, improving the mask and providing better performance for smaller objects, as shown in p3-p7 in Figure 3C.

[0197] In some embodiments, step B) comprises:

[0198] b1) performing feature extraction on the input image to obtain a feature backbone consisting of feature maps c1 to c5; and

[0199] b2) Obtain a Feature Pyramid Network (FPN) consisting of feature maps p3 to p7 from the obtained feature maps c1 to c5.

[0200] In some embodiments, in step b1), five feature maps c1 to c5 with sizes from large to small are generated by a fully convolutional network (FCN).

[0201] In some embodiments, in step b2), the feature map c5 with the smallest size is subjected to a convolution layer to obtain a feature map p5; the feature map p5 is amplified by a bilinear interpolation and added to the convolved feature map c4 to obtain a feature map p4; the feature map p4 is amplified by a bilinear interpolation and added to the convolved feature map c3 to obtain a feature map p3; the feature map p5 is convolved to obtain a feature map p6; the feature map p6 is convolved to obtain a feature map p7.

[0202] In some embodiments, step C) comprises:

[0203] c1) Send all feature maps p3 to p7 of the Feature Pyramid Network (FPN) to the detection branch to obtain the coefficient c+b+o for each anchor box, where the detection branch includes a prediction head and a non-maximum suppression module (NMS). The coefficient c represents the classification confidence, the coefficient b represents the bounding box regression, and the coefficient o represents the flower orientation.

[0204] c2) At the same time as step c1), the maximum feature map p3 of the feature pyramid network (FPN) is sent to the segmentation branch parallel to the detection branch, where m prototype masks are obtained.

[0205] In some embodiments, the flower orientation refers to the direction of the flower stigma, including the following five categories: left, right, front, top, and bottom.

[0206] YOLACT_Orient generates a "prediction head" from a sequence of feature maps. As shown in Figure 3C, the "prediction head" contains the target category (c) and target anchor box (b), where "c" represents classification information and "b" represents bounding box regression information. Simultaneously, Protonet outputs two mask images from feature map p7, indicating the maximum possible distinction between the presence of targets and non-target background regions in the image. The three types of information from the prediction head are combined with the results of Protonet through non-maximum suppression (NMS). First, the target location is determined by comparing the target category information ("c"), the target anchor box ("b"), and the mask image information. The target region is then segmented using a threshold. The flower anchor box is then used to crop the flower from the image, and each object is segmented pixel-wise within the cropped region. This process is illustrated in the "Crop" and "Threshold" modules in Figure 3C.

[0207] Furthermore, as shown in FIG3C , the “prediction head” generated by YOLACT_Orient from the feature map sequence not only includes the target category (c) and the target anchor box (b), but also includes the target direction information (o). The flower orientation described by the applicant in the experiment refers to the direction of the flower stigma, which includes the following five categories: left, right, front, top, and bottom, but the present invention is not limited to this. After obtaining the cropped target position information by comparing the prediction head with the mask image information, the target orientation information is further obtained by comparing the target orientation information (“o”) after maximum suppression with the cropped range. This process is shown in the NMS red line in FIG3C and the Orientation after the “Crop” and “Threshold” modules.

[0208] Regarding network design, the present invention also adds a network branch (Orient) to the loss function to distinguish the direction of the flower. The initial parameter vector (x, y, w, h, c, m) is then converted to (x, y, w, h, c, m, o), where o represents the direction of the flower stigma. Here, the flower orientation described by the applicant in the experiment refers to the direction of the flower stigma, which includes the following five categories: left, right, front, top and bottom, but the present invention is not limited to this. Figure 9A shows five examples of flower stigma directions, where the blue circle represents the center of the flower anchor frame and the red arrows emanating from the center of the flower represent the five directions. The green dotted line represents the orientation range of the orientation.

[0209] In some embodiments, the loss function is represented by the following formula (1): L = wcls ·L cls +w box ·L box +w mask ·L mask +w orient ·L orient (1)

[0210] in,

[0211] L cls Represents the classification confidence loss of whether the target category is a flower, and its weight coefficient is w cls ,

[0212] L box Represents the bounding box regression loss, whose weight coefficient is w box ,

[0213] L mask Represents the mask loss, and its weight coefficient is w mask ,

[0214] L orient Represents the flower orientation loss, and its weight coefficient is w orient .

[0215] In some embodiments, w cls Set to 1.0, w box Set to 1.5, w mask Set to 6.125, w orient The weight coefficient setting of the present invention is not limited to this, and can be adjusted according to the goals of different flowers in actual applications.

[0216] In some embodiments, the flower orientation loss Lorient is represented by the following formula (2):

[0217] Where θ represents the direction angle of the flower stigma. The range of θ is consistent with the range of mainstream target oblique frame detection methods, such as OBB (Objection Bouding Box), which is usually 180°. In the present invention, it can be set to

[0218] In some embodiments, the output result obtained in step D) is (x, y, w, h, c, m, o), where x, y, w, h represent the position, width, and height of the flower box, c represents the flower classification confidence, m represents the number of prototype masks, and o represents the flower's top, bottom, left, right, and front orientations. Ultimately, YOLACT_Orient outputs the target center position, the box position, width, and height, the pixels that constitute the target, and the orientation of the flower's stigma.

[0219] 6. Automated Plant Hybrid Breeding Methods

[0220] In one aspect, the present invention provides a method for automated plant hybrid breeding using an automatic pollination robot, the method comprising:

[0221] i) providing a plant having a flower with an exserted stigma phenotype as a hybrid mother plant;

[0222] ii) allowing the automatic pollination robot to approach the plant and acquire an image of the plant;

[0223] iii) enabling an automatic pollination robot to process the image using an artificial intelligence method, detect the flowers of the plant, and determine the three-dimensional position of the stigma; and

[0224] iv) causing an automatic pollination robot to pollinate the stigmas in the flowers of the hybrid female plant with pollen from the hybrid male plant.

[0225] In some embodiments, the method further comprises v) harvesting hybrid seeds from the hybrid female plant.

[0226] In some embodiments, the hybrid mother plant in step i) is the modified plant described above in the present invention or a modified plant obtained by the method described above in the present invention.

[0227] In some embodiments, the automatic pollination robot is as described above. In some embodiments, the automatic pollination robot may include a movable base; a robotic arm assembly mounted on the base; a pollination device, a camera, and a pollen box, each detachably attached to the robotic arm assembly; and a controller.

[0228] The base, robotic arm assembly and camera can be corresponding commercially available products, and the pollination device can be, for example, a wool pen or a wool brush, which can be grasped by the end of the robotic arm assembly, such as a robotic claw, to be detachably attached to the robotic arm assembly. The pollination device can be moved to the pollen box to obtain pollen and can be moved to the stigma of the crop to be pollinated by the robotic arm assembly.

[0229] In some embodiments, the automatic pollination robot further comprises a positioning module. In some embodiments, through the positioning module, for example, through the positioning function of LocalSense and a map-based scanning positioning strategy, the controller can locate and navigate the movable base to the target hybrid female plant to be pollinated.

[0230] In some embodiments, in step iii), the automated pollination robot processes the image using the method described above. In some embodiments, in step iii), the automated pollination robot processes the image using the following method to detect the flowers of the plant and determine the three-dimensional position of the stigma:

[0231] A) Input image containing flowers;

[0232] B) Send the input image to the backbone network for feature extraction to obtain a feature map;

[0233] C) sending the obtained feature map to the detection branch to obtain prediction coefficients, and sending the obtained feature map to the segmentation branch to obtain a prototype mask, wherein the obtained prediction coefficients include coefficients representing the flower orientation; and

[0234] D) The obtained prototype mask is combined with the prediction coefficients, and the output result is obtained together with the coefficient representing the flower orientation through the cropping and threshold modules.

[0235] In some embodiments, in step A), the image is pre-processed, wherein the size of the input image is adjusted to W×H×3. In some embodiments, the pixel resolution of W and H is set to 550. However, the pixel resolution setting of the present invention is not limited thereto.

[0236] Based on the YOLACT network framework (Bolya, D., et al., (2022). YOLACT++Better Real-Time Instance Segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(2), 1108-1121. DOI: 10.1109 / TPAMI.2020.3014297), the present invention adds a flower orientation classification network, which is called YOLACT_Orient in the present invention. YOLACT_Orient can not only simultaneously detect and segment the range of flowers, but also infer the orientation of flowers.

[0237] The feature main network marked as c1-c5 in Figure 3C is built by generating masks using Fully Convolutional Networks (FCN).

[0238] Higher resolution pyramid prototypes are achieved by using Feature Pyramid Networks (FPN) to connect c3 and p3 with the same size. FPN then continuously upsamples to a quarter of the input image size, improving the mask and providing better performance for smaller objects, as shown in p3-p7 in Figure 3C.

[0239] In some embodiments, step B) comprises:

[0240] b1) performing feature extraction on the input image to obtain a feature backbone consisting of feature maps c1 to c5; and

[0241] b2) Obtain a Feature Pyramid Network (FPN) consisting of feature maps p3 to p7 from the obtained feature maps c1 to c5.

[0242] In some embodiments, in step b1), five feature maps c1 to c5 with sizes from large to small are generated by a fully convolutional network (FCN).

[0243] In some embodiments, in step b2), the feature map c5 with the smallest size is subjected to a convolution layer to obtain a feature map p5; the feature map p5 is amplified by a bilinear interpolation and added to the convolved feature map c4 to obtain a feature map p4; the feature map p4 is amplified by a bilinear interpolation and added to the convolved feature map c3 to obtain a feature map p3; the feature map p5 is convolved to obtain a feature map p6; the feature map p6 is convolved to obtain a feature map p7.

[0244] In some embodiments, step C) comprises:

[0245] c1) Send all feature maps p3 to p7 of the Feature Pyramid Network (FPN) to the detection branch to obtain the coefficient c+b+o for each anchor box, where the detection branch includes a prediction head and a non-maximum suppression module (NMS). The coefficient c represents the classification confidence, the coefficient b represents the bounding box regression, and the coefficient o represents the flower orientation.

[0246] c2) At the same time as step c1), the maximum feature map p3 of the feature pyramid network (FPN) is sent to the segmentation branch parallel to the detection branch, where m prototype masks are obtained.

[0247] In some embodiments, the flower orientation refers to the direction of the flower stigma, including the following five categories: left, right, front, top, and bottom.

[0248] YOLACT_Orient generates a "prediction head" from a sequence of feature maps. As shown in Figure 3C, the "prediction head" contains the target category (c) and target anchor box (b), where "c" represents classification information and "b" represents bounding box regression information. Simultaneously, Protonet outputs two mask images from feature map p7, indicating the maximum possible distinction between the presence of targets and non-target background regions in the image. The three types of information from the prediction head are combined with the results of Protonet through non-maximum suppression (NMS). First, the target location is determined by comparing the target category information ("c"), the target anchor box ("b"), and the mask image information. The target region is then segmented using a threshold. The flower anchor box is then used to crop the flower from the image, and each object is segmented pixel-wise within the cropped region. This process is illustrated in the "Crop" and "Threshold" modules in Figure 3C.

[0249] Furthermore, as shown in FIG3C , the “prediction head” generated by YOLACT_Orient from the feature map sequence not only includes the target category (c) and the target anchor box (b), but also includes the target direction information (o). The flower orientation described by the applicant in the experiment refers to the direction of the flower stigma, which includes the following five categories: left, right, front, top, and bottom, but the present invention is not limited to this. After obtaining the cropped target position information by comparing the prediction head with the mask image information, the target orientation information is further obtained by comparing the target orientation information (“o”) after maximum suppression with the cropped range. This process is shown in the NMS red line in FIG3C and the Orientation after the “Crop” and “Threshold” modules.

[0250] Regarding network design, the present invention also adds a network branch (Orient) to the loss function to distinguish the direction of the flower. The initial parameter vector (x, y, w, h, c, m) is then converted to (x, y, w, h, c, m, o), where o represents the direction of the flower stigma. Here, the flower orientation described by the applicant in the experiment refers to the direction of the flower stigma, which includes the following five categories: left, right, front, top and bottom, but the present invention is not limited to this. Figure 9A shows five examples of flower stigma directions, where the blue circle represents the center of the flower anchor frame and the red arrows emanating from the center of the flower represent the five directions. The green dotted line represents the orientation range of the orientation.

[0251] In some embodiments, the loss function is represented by the following formula (1): L = wcls ·L cls +w box ·L box +w mask ·L mask +w orient ·L orient (1)

[0252] in,

[0253] L cls Represents the classification confidence loss of whether the target category is a flower, and its weight coefficient is w cls ,

[0254] L box Represents the bounding box regression loss, whose weight coefficient is w box ,

[0255] L mask Represents the mask loss, and its weight coefficient is w mask ,

[0256] L orient Represents the flower orientation loss, and its weight coefficient is w orient .

[0257] In some embodiments, w cls Set to 1.0, w box Set to 1.5, w mask Set to 6.125, w orient The weight coefficient setting of the present invention is not limited to this, and can be adjusted according to the goals of different flowers in actual applications.

[0258] In some embodiments, the flower orientation loss Lorient is represented by the following formula (2):

[0259] Where θ represents the direction angle of the flower stigma frame. The range of θ is consistent with the range of mainstream target oblique frame detection methods, such as OBB (Objection Bouding Box), which is usually 180°. The value of θ in this patent is set to Of course, any continuous 180° on the circumference can be set as the calculation range of θ.

[0260] In some embodiments, the output result obtained in step D) is (x, y, w, h, c, m, o), where x, y, w, h represent the position, width, and height of the flower box, c represents the flower classification confidence, m represents the number of prototype masks, and o represents the top, bottom, left, right, and front directions of the flower stigma. Ultimately, YOLACT_Orient outputs the target center position, the box position, width, and height, the pixels that constitute the target, and the orientation of the flower stigma.

[0261] In some embodiments, step iii) further comprises positioning the stigma using a "pseudo-binocular" method.

[0262] Using binocular vision for 3D positioning is well known in the art. It typically involves simultaneously capturing an object with two cameras at different locations, and then determining the 3D position of the object relative to the cameras based on the matching of feature points between the two images. However, in the field of automated plant pollination, directly attaching two cameras to the end of a robotic system increases the system's calibration costs and increases the likelihood that the cameras and other accessories will scratch or even damage plants, particularly stigmas, during system movement.

[0263] To overcome the above shortcomings, the present invention adopts a "pseudo-binocular" three-dimensional positioning method. Specifically, the automatic pollination robot in the present invention is equipped with only one camera. Through precise manipulation at the end of the robotic arm, the camera moves to obtain images of the stigma at two different positions, and obtains the three-dimensional precise positioning of the stigma based on the matching of feature points between the images. Therefore, compared with the binocular system using two cameras, the "pseudo-binocular" three-dimensional positioning method of the present invention achieves at least equivalent positioning accuracy, while also reducing the calibration cost and space occupied by the system, and reducing the possibility of scratching or even damaging the plants, especially the stigma.

[0264] The camera may photograph the column head at one position and then move horizontally to another nearby position to photograph the column head. The distance between the two positions is, for example, 10 cm, but the present invention is not limited thereto.

[0265] Image feature point matching algorithms are well known in the field of image processing. For example, feature point matching algorithms such as SIFT (Scale-invariant feature transform) and SURF (Speeded Up Robust Features) can be combined with the RANSAC (Random Sample Consensus) algorithm to obtain pre-matched feature points between pseudo-binocular image pairs. However, the present invention is not limited to this, and other feature point matching algorithms can also be used to obtain the three-dimensional location of the column head. It should be noted that the three-dimensional location obtained in this case is the location (three-dimensional coordinates) of the column head in the camera coordinate system. Specifically, the three-dimensional location can be the three-dimensional coordinates of the center of the column head in the camera coordinate system.

[0266] When matching feature points, other parts of the plant besides the petals and objects in the crop's surroundings may constitute background noise and adversely affect the accuracy of the matching, thereby reducing the three-dimensional positioning accuracy of the stigma. To this end, an image of the petal area including the stigma can be cropped from the image acquired by the camera, and feature point matching can be performed only based on the cropped image, thereby removing background noise and improving the three-dimensional positioning accuracy of the stigma. Moreover, since the size of the cropped image becomes smaller, the speed of calculating the three-dimensional positioning of the stigma will also be accelerated. According to the applicant's experiments, the average time required to calculate the three-dimensional positioning of the stigma is approximately 0.045 seconds. In this way, the stigma can be quickly positioned in three dimensions.

[0267] In some embodiments, in step iv), the controller drives the pollination device and the pollen box carried by the robotic arm assembly to move in front of the plant stigma based on the three-dimensional positioning of the stigma.

[0268] At this time, it is necessary to convert the three-dimensional positioning of the stigma in the camera coordinate system into its three-dimensional positioning in the system coordinate system of the automatic pollination robot. As long as the camera is calibrated after being installed on the robot, the specific position of the camera in the system can be known, or the three-dimensional positioning of the camera relative to the pollination device of the robot (such as a pollination brush) can be known. Therefore, as long as the three-dimensional positioning of the stigma relative to the camera is obtained, it is equivalent to obtaining the three-dimensional positioning of the stigma relative to the pollination device. In order to complete the subsequent pollination operation, it is necessary to first move the pollination device along the stigma toward the corresponding direction to the vicinity of the stigma to be pollinated. For example, based on the three-dimensional positioning of the stigma relative to the pollination device, the pollination device is driven to move to the front near the stigma, for example, 10 cm directly in front of the stigma, but the present invention is not limited to this.

[0269] The positioning algorithm based on feature point matching may have certain positioning errors. Taking the applicant's experiment of three-dimensional positioning of the stigmas of 64 randomly selected tomato flowers as an example, by comparing the three-dimensional positioning of the stigmas obtained by the "pseudo-binocular" three-dimensional positioning method with the three-dimensional positioning of the stigmas actually measured, it was found that the stigmas had an average positioning error of 0.50cm, 0.85cm and 0.53cm in the three axes of the three-dimensional coordinate system X, Y, and Z, respectively, and a maximum positioning error of 1.33cm, 2.22cm and 1.29cm in the three axes, respectively. Due to the existence of positioning errors, if the pollination device is directly driven to move to the three-dimensional positioning of the stigma obtained by the positioning algorithm based on feature point matching, the pollination device may not contact the stigma, resulting in unsuccessful pollination.

[0270] In order to increase the success rate of pollination, in step iv) of the method of the present invention, after the pollination device moves to the front near the stigma, the controller drives the pollination device carrying pollen to move sequentially between multiple points in a set three-dimensional space, including the three-dimensional positioning of the stigma obtained according to the previous steps, until the pollination device coats the pollen onto the stigma.

[0271] That is, the method according to the present invention adopts a "coarse to fine" strategy to achieve pollination, first moving the pollination device to the vicinity of the stigma, and then moving it through multiple points to achieve a fine search of the stigma.

[0272] The aforementioned three-dimensional space can be understood as the spatial range for the pollination device to retrieve the stigma, which can be set as a regular three-dimensional space, such as a sphere, a cuboid, a cube, a cylinder, etc., or it can also be set as an irregular three-dimensional space. Preferably, the three-dimensional space is set as a three-dimensional space that covers the maximum positioning error of the three-dimensional positioning of the stigma obtained by matching the feature points in the three-dimensional coordinate system X, Y, and Z. Taking the maximum positioning errors of 1.33 cm, 2.22 cm, and 1.29 cm in the three-dimensional coordinate system X, Y, and Z obtained by the applicant's experiment as an example, the above-mentioned three-dimensional space can be set as a three-dimensional space formed by taking the three-dimensional positioning of the stigma obtained by matching the feature points as the center, extending 1.33 cm in the positive and negative directions of the X axis, 2.22 cm in the positive and negative directions of the Y axis, and 1.29 cm in the positive and negative directions of the Z axis.

[0273] Therefore, as long as the pollination device is driven to move sequentially between enough points in the aforementioned three-dimensional space, it can be ensured that when the pollination device moves to a certain point, the pollen will be successfully applied to the stigma to complete pollination. For example, the multiple points in the aforementioned three-dimensional space are selected to be evenly distributed in the three-dimensional space, and the number of points can be determined based on the error size of the three-dimensional positioning of the stigma obtained by matching the feature points. In other words, the number of points is adjustable to improve the pollination efficiency while ensuring the success rate of pollination. Moreover, once successful pollination is determined, the pollination device is stopped from being driven to continue moving to other points in the three-dimensional space. Instead, the existence of other crop stigmas to be pollinated is re-detected, and the pollination device is driven to move to the vicinity of the newly detected crop stigma and start new pollination. Thereby, the pollination efficiency is further improved.

[0274] Because the stigma is more likely to be located at or near the 3D location of the stigma obtained through feature point matching, and less likely to be located at the outer contour of the 3D space, the pollination device can be driven to first move to the stigma's 3D location obtained through feature point matching, and then gradually move to the outer contour of the 3D space through multiple points from near to far. Once successful pollination is confirmed, the pollination device is stopped and continues to move to other points in the 3D space. This further improves pollination efficiency.

[0275] The applicant has discovered that if the aforementioned three-dimensional space is set as an irregular solid space, selecting the point within the space to which the pollination device is to move becomes relatively complex, and it is also not conducive to path planning for the pollination device to move sequentially between multiple points. Therefore, it is preferable to set the three-dimensional space as a regular solid space.

[0276] Specifically, the present invention has found that it is particularly advantageous to construct the aforementioned three-dimensional space into a cylinder, which covers the maximum positioning error of the three-dimensional positioning of the column head obtained by matching feature points in the three-dimensional coordinate system X, Y, Z.

[0277] To facilitate the retrieval of the stigma, the pollination device is driven to sequentially move along multiple points along the circular spiral lines in at least two circular cross-sections of the cylinder until the stigma is pollinated. In other words, the pollination device sequentially moves along multiple points along at least two parallel circular spiral lines in the three-dimensional space of the cylinder until successful pollination is determined. In this manner, the pollination device only needs to translate to the position of each circular spiral line and then sequentially move along multiple points along a circular spiral line in a vertical plane at that position to ensure that the pollination device successfully applies pollen to the stigma during the process.

[0278] In order to determine whether the pollination device has successfully applied pollen to the stigma, the method adopted is to use a camera to obtain an image containing the stigma and the pollination device every time the pollination device moves to a point, and determine whether the two are in contact based on their relative positions in the image. It should be noted that since this application only uses one camera, there may be a possibility of misjudgment when judging whether pollination is successful based on only one image obtained by it. This misjudgment may be caused by factors such as interference from other objects around the stigma. Therefore, in order to improve the accuracy of judging whether pollination is successful, the method of the present invention proposes to determine that pollination is successful only when the pollination device and the stigma are judged to be in contact twice in a row.

[0279] To expedite the process of determining whether the pollination device has successfully applied pollen to the stigma, the present invention utilizes a lightweight network architecture based on Inception-v3. Each time an image containing the stigma and the pollination device is acquired, it can quickly confirm whether the stigma and the pollination device are in contact. The Inception-v3 network architecture is well known in the art and will not be described in detail here. Of course, any visual-tactile processing method for detecting whether the stigma has been touched by the pollination device is applicable to the present invention.

[0280] In some embodiments, the plant is a Solanaceae plant. In some embodiments, the plant is selected from Solanum lycopersicum (tomato), Nicotiana benthamiana (tobacco), Capsicum annuum (pepper), Physalis pruinosa (mushroom), Solanum melongena (eggplant), Solanum tuberosum (potato), Solanum pennellii, Solanum chilense, Solanum habrochaites, Solanum pimpinellifolium, Solanum galapagense, and Petunia hybrid (petunia). Preferably, the plant is Solanum lycopersicum (tomato).

[0281] In some embodiments, the plant can be from a different cultivar of Solanum lycopersicum (tomato). In some embodiments, the plant is tomato cultivar Ailsa Craig, M82, Beijing 1, TS545, TS181, TS590.

[0282] In some embodiments, the plant is a leguminous plant such as soybean.

[0283] In some embodiments, the hybrid mother plant is a modified plant having an exserted stigma phenotype as described above. For example, the modified plant is a modified Solanum lycopersicum (tomato) or soybean. In some specific embodiments, the modified plant is from the tomato cultivar Ailsa Craig.

[0284] In some embodiments, the hybrid male plant is a plant that can produce fertile pollen and can pollinate the hybrid female plant.In some embodiments, the hybrid male plant is a plant with excellent agronomic traits.

[0285] For example, the hybrid male plant is Solanum lycopersicum (tomato). In some specific embodiments, the modified plant is from tomato cultivars M82, Beijing1, TS545, TS181, TS590.

[0286] In some embodiments, the hybrid male plant is wild tomato Solanum pimpinellifolium or a progeny thereof, such as a progeny thereof having superior agronomic traits.

[0287] In some embodiments, the hybrid male and / or female plants are cultured in an LED-based phytotron. In some embodiments, the hybrid male and / or female plants are cultured under LED lighting and a day length of approximately 22 hours.

[0288] In one aspect, the present invention provides a method for tomato hybrid breeding, comprising:

[0289] i) providing the tomato cultivar Ailsa Craig as a hybrid parent plant;

[0290] ii) providing tomato cultivar M82, Beijing1, TS545, TS181 or TS590 as the hybrid male parent plant;

[0291] iii) pollinating the stigmas in the flowers of the female hybrid plant with pollen from the male hybrid plant.

[0292] In some embodiments, the method further comprises iv) harvesting hybrid seeds from the hybrid female plant. Example

[0293] The present invention can be further understood by reference to the specific embodiments described herein, which are intended to illustrate the present invention only and are not intended to limit the scope of the present invention. Obviously, those skilled in the art will appreciate that various modifications and variations can be made to the present invention without departing from the spirit of the present invention, and such modifications and variations also fall within the scope of the present invention.

[0294] Experimental Materials and Methods

[0295] Plant materials and growth conditions

[0296] Seeds of the cultivated tomato variety Ailsa Craig and the wild species Solanum pimpinellifolium were obtained from the Tomato Genetic Resource Center (https: / / tgrc.ucdavis.edu / ). Seeds of M82, Beijing1, and three heirloom accessions (TS545, TS181, and TS590) were provided by the laboratories of Z. Lippman, S. Wang, and S. Huang, respectively. The glo2 (Solyc06g059970) mutant generated in this study and the newly domesticated wild eggplant (Solanum pimpinellifolium sp sp5g) were propagated and maintained in our laboratory. Male sterile mutants were maintained as heterozygous seeds. To collect representative flower images, the selected Solanaceae species were grown in a laboratory greenhouse.

[0297] All plant materials were grown in Beijing, China. Seeds were sown directly into soil in 72-well plastic plug trays and grown in a natural light greenhouse supplemented with LED lights (Philips Lighting IBRS, 10461, 5600VB, NL) with a 14 h L / 10 h D photoperiod, or in an artificial climate chamber equipped with LED lights (Philips Lighting IBRS, 10461, 5600VB, NL) and air conditioning. The conditions in the artificial climate chamber included a photoperiod (16 h L / 8 h D), a light intensity (120 μmol·m -2 ·s -1 ), daytime and nighttime temperatures (26-28°C and 18-20°C, respectively), and humidity (45-60%). Unless otherwise specified, agronomic trait analyses were performed under greenhouse (March-July, August-December) or open-air cultivation (March-July) conditions.

[0298] gRNA design and CRISPR-Cas9 knockout vector construction

[0299] The Csy4-multi-gRNA CRISPR / Cas9 system was expressed using the binary vector pDIRECT_22C. Csy4 (187 amino acids) was linked to the Arabidopsis codon-optimized SpCas9 via P2A, and expression was driven by the cauliflower mosaic virus (CaMV) 35S promoter. The main steps included target site design (http: / / crispr.hzau.edu.cn / CRISPR2 / ), primer design (http: / / crispr-multiplex.cbs.umn.edu / ), and vector construction. Using Golden Gate cloning, gRNAs were isolated from the 20-bp csy4 binding site and introduced into pDIRECT_22C all at once. For GLO2, gRNAs were designed to target the coding region, noncoding region, and the second intron as follows: target-K1 and target-K2 targeted exons 2 and 7, respectively; target-2, target-3, target-5, and target-6 targeted the nonprotein coding region; and target-1-4 targeted the second intron. The three gene editing vectors were constructed separately. The tomato gene sequence was obtained from the Solanaceae genome website (https: / / solgenomics.net / ).

[0300] Generation of CRISPR lines and genotyping

[0301] Following previous studies, genetic transformation was performed using Agrobacterium tumefaciens LBA4404. Successfully rooted T0 plants were selected for genotyping. Leaf samples were collected from three different locations within each line to identify all possible mutations, and genomic DNA was extracted using the CTAB method. The mutation patterns of the edited T0 and T1 plants were confirmed by PCR amplification of a fragment containing the target gene, followed by T7E1 digestion, Sanger sequencing, and massively parallel sequencing.

[0302] Sequencing analysis

[0303] The Illumina NovaSeq platform was used for massively parallel sequencing to determine the genotype of the glo2-in2-a2 mutation carrying the structural variation. To extract high molecular weight DNA, the inventors collected the sixth true leaf during the vegetative growth phase of glo2-in2-a2. DNA extraction, library preparation, and sequencing were performed by Berry Biosciences. Paired-end reads (150 bp) were sequenced using the Illumina NovaSeq platform, with an insert length of approximately 350 bp. A total of 64,660,723 reads were obtained, and the raw data were quality-processed to obtain clean data. For data analysis, the resequenced FASTQ files were aligned to the reference genome SL4.0 using bwa-mem (version 0.7.17) software. Pindel (version 0.2.5a3) software was used for larger inserts and structural alignment. After analyzing the data, the sequence information of glo2-in2-a2 was obtained and its mutation form was determined.

[0304] Microscopic imaging

[0305] Tomato flowers with fully expanded petals were used for analysis. Petals and stamens were removed sequentially using a dissecting needle and forceps. The ovary was then dissected open using an ultra-thin blade. This procedure was performed using a Zeiss SteREO Discov V20@1126 microscope equipped with Z-Stack acquisition. Optical sections were aligned and merged using AxioVision Rel. 4.8 software (Zeiss) to produce the final focused image.

[0306] Pollen collection for robotic pollination

[0307] Select mature tomato flowers with fully expanded petals. Carefully separate the stamens with a dissecting needle, place on sintered paper, and dry in a 28°C oven for 10 hours. Grind the dried stamens and sieve through a 70 μm cell sieve (LANSO, Cat# LS-GA8002) to remove debris. Collect the pollen. Finally, aliquot the dried pollen into 1.5 mL tubes and store at -80°C until ready for use.

[0308] Pollen staining

[0309] Tomato pollen staining was performed according to the Alexander method with slight modifications. 20 μL of ddH2O was added to the collected pollen and vortexed to suspend it. 10 μL of the pollen solution was taken and 5 μL of Alexander staining solution (Solarbio, Cat#G3050) was added for staining. The resulting mixture was pipetted and mixed thoroughly, then aspirated onto a glass slide and vacuumed at 0.8 psi for 2 h. After vacuuming, the slide was incubated at room temperature overnight. Pollen staining was observed under bright field using a LEICA DM2500 LED microscope.

[0310] RNA extraction and qRT-PCR analysis

[0311] RNA was extracted from freshly collected, unopened flower buds (0.5–1 cm in length) using TRIzol (Invitrogen, Cat#15596018). The extracted RNA was treated with RNase Free DNase I (Invitrogen, Cat#AM2222) to remove genomic DNA. One μg of total RNA was reverse transcribed into cDNA using the FastKing cDNA First-Strand Synthesis Kit (TIANGEN, Cat#KR116-02). Gene-specific primers were designed for qRT-PCR analysis, using the tomato Ubiquitin (Solyc01g056940) gene as an internal control. qRT-PCR reactions were performed using TB Green Premix Ex Taq II (TaKaRa, Cat#RR820) and run on a CFX96 Real-time Detection System (Bio-Rad). Genes used for qRT-PCR analysis included SlGLO2, SlGLO1 (Solyc08g067230), SlDEF (Solyc04g081000), and TM6 (Solyc02g084630).

[0312] Phylogenetic analysis

[0313] Species phylogenetic analysis was performed using TimeTree (http: / / www.timetree.org / ). Protein sequences of class B MADS-box gene family members from representative Solanaceae species were obtained from the Solanaceae Genome Database (https: / / solgenomics.net / ), the Chinese Academy of Agricultural Sciences Tomato Database (http: / / caastomato.biocloud.net / home), or the National Genomics Data Center (https: / / ngdc.cncb.ac.cn / search / ?dbId=gwh&q=%20PRJCA010759). Sequences were aligned using ClustalW, and phylogenetic trees were constructed using the maximum likelihood (ML) method using MEGA7.

[0314] Determination of seed germination rate and 100-grain weight

[0315] Seed germination rate was determined by placing seeds on moistened filter paper and germinating at room temperature in the dark. All seeds were harvested and sterilized during the same growing season. Germination was measured every 24 hours for five times. 100-grain weight was measured using an ultramicrobalance. 100 seeds were randomly selected from each group. Eight biological replicates were performed for each genotype.

[0316] Robot hardware components

[0317] The pollination robot hardware shown in Figure 4A consists of seven components. First is the "carrier," a ShiHe MR1000 mobile robot platform that serves as the base for transporting the pollination robot to nearby plant areas within the greenhouse. It has a maximum load capacity of 200 kg and a maximum speed of 2 m / s.

[0318] The second component is the Ultra-Wide Band (UWB) positioning module. Its main function is to accurately locate the two-dimensional (2D) position of the carrier in a specific greenhouse. To be precise, one UWB LocalSense gateway is installed at each of the four corners of a 12m×9.8m greenhouse area, and each gateway is 2.9m from the ground. By accurately measuring the transmission time of the wireless pulse, LocalSense can evaluate the absolute distance between the UWB gateway and the tag. In this way, the tag position can be calculated in real time at a frequency of more than 500Hz with an accuracy of up to 20cm. In addition, by using the STL (Scan-to-Locality) map navigation strategy, the pollination robot can locate and navigate to any target tomato plant within a range of 10cm (Figure 4E, F).

[0319] In this work, the inventors used the Universal Robot 5 (UR5) as a pollination arm. The UR5 has 6 degrees of freedom, an 85 cm working radius, and a payload capacity of 5 kg. Meanwhile, the fourth component in Figure 4A shows a Robotiq F85 Gripper attached to the end of the UR5 arm. The Robotiq F85 Gripper acts as a robotic arm, gripping the pollen brush and contacting the stigma.

[0320] The fifth component in Figure 4A is the RealSense D435i camera mounted on the top of the UR5. Its function is to identify flowers, distinguish stigmas, and provide visual data for the servo.

[0321] In this study, the pollen brush used was a soft wool pen in the shape of a 5 cm x 1 cm cylinder with a 1.5 cm tip. The pollen brush was gripped by a Robotiq F85 Gripper and moved synchronously with the UR5's end-of-line tooling to contact the stigma. The seventh component was the pollen container. This small cylindrical box, 2.5 cm in diameter and 1.5 cm high, was located at the front end of the ShiHe MR1000 mobile robot platform. During pollination, the robot must keep the pollen on the brush surface and frequently dip the brush head into the container to collect pollen for pollination.

[0322] The location of plants in the greenhouse and the robot's navigation route

[0323] Tomato plants were systematically placed in the greenhouse, as shown in the schematic diagram of the vertical layout of the greenhouse (Figure 4D). The spacing between tomato plants ranged from 0.4m to 0.6m, while mature plants bearing fruit were approximately 0.55m to 1.2m tall. In this study, the inventors determined that the distance between plants was 0.6m and the robot channel width was 2m. It is worth noting that this width matches the safe turning radius of the ShiHe MR1000 vehicle platform. In addition, the inventors used a small UWB signal marker on the ShiHe MR1000 mobile robot (Figure 4A), which enabled the inventors to obtain the position and orientation of the robot at a frequency of 200Hz. At the same time, using the positioning technology proposed by the LocalSense wireless positioning system determined in this study, combined with the STL (Scan-to-Locality) map navigation strategy, the vehicle can accurately transport the robotic arm to the plants in the greenhouse with a positioning error of no more than 10cm.

[0324] Stigma detection and orientation prediction based on deep learning

[0325] The inventors' goal is to accurately obtain the position of the tomato stigma and its orientation in the observed image. To this end, based on the widely recognized real-time YOLACT++ network structure and its high performance in object detection and segmentation, the inventors added a network branch (Orient) to its loss function to distinguish the direction of the flower. The inventors named this tomato flower detection, segmentation and stigma orientation prediction model YOLACT_Orient. Figure 3C shows the YOLACT_Orient structure proposed by the inventors, where the red part (arrows, lines and blocks) is the orientation prediction module proposed in the inventors' work, and the black part is the original subnetwork introduced in the original YOLACT structure. In Figure 3C, from left to right are A: input image; B: YOLACT main structure; C: original YOLACT loss function; D: proposed stigma orientation loss function; E: output result. Before entering YOLACT_Orient, the input image is resized to W×H×3. In this work, the pixel resolution ratio of W and H is set to 550. The output parameters (x, y, w, h, c, m, o) represent the confidence that the current box is a tomato flower (c), the box range of the tomato flower (x, y, w, h), the m prototype masks for instance segmentation, and the direction of the stigma (o).

[0326] (1)YOLACT_Orient main network

[0327] This work adopts the basic framework of YOLACT as the main body. The feature main network marked as c1-c5 in Figure 3C is established by generating masks using Fully Convolutional Networks (FCN).

[0328] Higher resolution pyramid prototypes are achieved by using Feature Pyramid Networks (FPN) to connect c3 and p3 with the same size. FPN then continuously upsamples to a quarter of the input image size, improving the mask and providing better performance for smaller objects, as shown in p3-p7 in Figure 3C.

[0329] (2) Tomato Flower Detection and Segmentation Sub-Network

[0330] A prediction head is generated for each anchor point in YOLACT_Orient. The "Prediction Head" column in Figure 3C shows the coefficients "c+b+o" for each anchor point, where the "c" term represents the classification confidence, the "b" term represents the bounding box suppression factor, and the "o" (orient) term represents the orientation prediction, which was proposed by the present inventors for orientation detection. The present inventors will introduce "o" in the next paragraph. Thus, the flower anchor point in the image can be obtained from each "b+c" term in the prediction head. Simultaneously, FCN and FPN always generate s prototype masks for the entire image. The prototype masks are listed in the upper box of the prototype module. The flower anchor point box can then be used to crop the flower from the image, and each object can be segmented pixel-wise from the cropped area. Segmentation is performed using a binary threshold combined with the prototype masks. The tomato flower segmentation process is shown in the "Crop" and "Threshold" modules in Figure 3C.

[0331] (3) Determine the direction of the tomato stigma

[0332] The inventors added a network branch (Orient) to the YOLACT loss function to distinguish the direction of the flower. The initial parameter vector (x, y, w, h, c, m) of YOLACT is then converted to (x, y, w, h, c, m, o), where o represents the direction of the stigma. A similar strategy has been proposed to classify the directions of flowers into three categories: "left", "front" and "right". Here, the inventors classify the stigma directions into five categories: "left", "right", "front", "up" and "down". Figure 9A shows five examples of flower stigma directions, where the blue circle represents the center of the flower anchor box and the red arrows emanating from the center of the flower represent the five directions. The green dotted line represents the orientation range of the orientation.

[0333] (4) Loss function design

[0334] With the definitions of flower anchor boxes, classification, orientation, and mask, the inventors calculated the loss function of the proposed YOLACT_Orient as follows. Let L cls 、L box 、L mask are the original loss functions for tomato flower classification, anchor range regressor, and mask coefficient respectively. The loss function of the proposed "orient" network is denoted as L orient Thus, the new loss function of the proposed YOLACT_Orient is given by (1), where w cls 、w box 、w mask and w orient L cls 、L box 、L mask and L orient According to the YOLACT approach, the inventors will w cls 、w box and w mask The values ​​of are set to 1.0, 1.5 and 6.125 respectively. orient Set to 1.0. YOLACT_Orient =w cls ·L cls +w box ·L box +w mask ·L mask +w orient ·L orient (1)

[0335] In (1), L orient It is treated as a categorical parameter and encoded with a single feature vector containing 5 items. orient The value of is represented by (2). In (2), θ represents the direction angle of the flower stigma frame, and its range is consistent with the range of mainstream target oblique frame detection methods, such as OBB (Objection Bouding Box), which is usually 180°. The value of θ in this patent is set to Of course, any continuous 180° on the circumference can be set as the calculation range of θ. The joint calculation of θ and O helps to distinguish "front", "left", "right", "up", and "down" during the pre-training annotation process.

[0336] Flower stigma pose estimation

[0337] In order to accurately deliver pollen to flowers, the inventors need to obtain the position of the stigma in the robotic arm coordinate system. In order to obtain the position of the stigma, the inventors first used a pseudo-binocular ranging strategy, combined with feature point matching (Speeded Up Robust Features, SURF) and RANSAC (Random Sample Consensus), to calculate the three-dimensional position of the stigma in the eye-hand camera coordinate system. Then, the stigma position in the camera coordinate system is converted to the robot arm coordinate system through eye-hand calibration. The pseudo-binocular ranging strategy configuration is shown in Figure 3F, where C1 (x1, y1, z1) is the position of the current camera in the robotic arm operation coordinate system. When a flower is found in the image taken at C1, the system records the position of C1 and moves the camera to a new position C2 (x2, y2, z2) = C1 + (Δx, 0, 0), where Δx = (x2-x1), which is the moving distance in the horizontal direction along the x-axis. In this article, Δx is set to 10 cm based on experience. In Figure 3F, X C (x C ,y C , z C ) and X R (x R ,y R , z R ) are the three-dimensional positions of the identified flower center in the camera coordinates C1 and the robot arm operation coordinates, respectively. I1 = (u1, v1) and I2 = (u2, v2) are the center points of the pistil region in the images captured by the camera at C1 and C2, respectively. The inventors used SURF and RANSAC strategies to obtain the flower center point by matching feature points on I1 and I2 observed at C1 and C2, respectively (Figure 3F). Then, using (3), (4) and (5), we can get X C , where f x , f y are the focal lengths of the camera along the x-axis and y-axis respectively, and u0 and v0 are the image centers measured in pixels. C =(Δx*f x ) / (u2-u1) (3) x C =(u1-u0)z C / f x (4) y C =(v1-v0)z C / f y (5)

[0338] Get X C Then, according to (6), the robot arm coordinate X can be obtained R The center of the flower.R =K[R|T]X C (6)

[0339] In (6), K is the camera intrinsic parameter matrix, which can be obtained by camera intrinsic parameter calibration. R and T are the rotation matrix and translation matrix from the camera coordinate system to the robot arm motion coordinate system, respectively. The inventors used the eye-hand calibration method to calculate R and T. The inventors calculated the average, maximum, minimum and variance range between the calculated and actual stigma positions measured from 64 randomly selected tomato flowers. The average positioning error of the stigma position was 0.50 cm (maximum distance 1.33 cm), 0.85 cm (maximum distance 2.22 cm) and 0.53 cm (maximum distance 1.29 cm) in the X, Y and Z dimensions, respectively (Figures 3G and 9C). The inventors only processed the part of the image cropped from the petal area by two-dimensional detection. This enables feature matching and RANSAC to resist background noise and ensures calculation speed (the average time for calculating the three-dimensional position of the stigma is 0.045 s) (Figure 3G). Therefore, the average positioning error of 0.50–0.85 cm and the average time of 0.045 s to calculate the 3D position of the stigma may help the robotic arm to quickly move the pollen brush toward the nearby stigma.

[0340] Servo-based stigma contact with the pollen brush

[0341] After determining the position and orientation of the pistil, the robot executes a coarse-to-fine servo pollination strategy. This strategy consists of two steps: "Flower Reaching" and "Servoing," corresponding to the coarse and fine phases, respectively. The coarse step aims to move the pollen brush to a position near the flower. Based on the position of the pistil, the "Flower Reaching" step can be executed using an inverse kinematics strategy.

[0342] The inventors adopted a circular search trajectory servo strategy to control the pollen brush to cover the stigma. Figure 4G shows the details of the circular search trajectory servo strategy, where the 12 small arrow positions along the circular spiral are the positions of the pollen brush servo. The symbol g is the span of the circular spiral. The symbol d is the diameter of the pollen brush. The servo strategy in the robot arm operation space is shown in the right column of Figure 4G. The s (1≤s≤S) circular search trajectory is arranged along an axis orthogonal to the flower surface, and the distance between each circle is d. Then, the servo space is shaped into a cylinder with a maximum volume of π×Sd×(g+d) 2 / 4. In the experiment, S is set to 3.

[0343] To ensure safe delivery of pollen to the stigma, the end effector uses a pollen brush with a 1.5 cm wool end. It then moves along an axis perpendicular to the flower until it lightly contacts the stigma surface. Since the error in identifying the stigma position is approximately 1.3-2.2 cm, the inventors further utilized a lightweight network architecture, Inception-v3, to classify the stigma as either contacted or uncontacted from the camera view. Within an acceptable time budget, the pollination brush is guided by a maximum of k = 36 servo points. At the same time, if the stigma contact based on Inception-v3 is confirmed, the system will stop servoing the current stigma. In 320 experiments, the pollination success rates of k = 23 and k = 29 servo points were 94.45% and 99.68%, respectively (Figure 4H, Figure 4I).

[0344] Robotic pollination process

[0345] The inventors used the aforementioned programs and tools to create an autonomous pollination robot. All models were deployed on a portable laptop computer equipped with a 2.60GHz CPU, 8.0GB of RAM, and an NVIDIA RTX 2080Ti GPU. The robotic pollination process in a greenhouse is shown in Figure 9G. The process consists of four stages:

[0346] During this phase, the carrier aims to transport the pollination robot to a location near the plants in the greenhouse. The ShiHe MR1000 mobile robot platform, combined with a UWB positioning module and an STL map navigation strategy, provides positioning. The pollination robot can detect and navigate to any desired tomato plant within a 10 cm range (Figure 4E, F).

[0347] Stigma perception. After the pollination robot arrives near the target plant, the system begins identifying the flower and its stigma. In this step, the inventors follow the strategy described in the "Deep Learning-Based Stigma Detection and Orientation Prediction" section above to obtain the position and orientation of the tomato stigma in the observed image. Using the proposed deep learning architecture, YOLACT_Orient, the system was able to identify the position and orientation of the stigma in just 0.12 seconds across 320 experimental runs.

[0348] Flower stigma pose estimation. This stage aims to determine the position of the flower stigma in the robot arm's coordinate system. The inventors described stigma pose estimation in the "Flower Stigma Pose Estimation" section above. By combining a pseudo-binocular ranging strategy with feature point matching (SURF) and RANSAC, the system achieves an average positioning error of 0.50-0.85 cm, with an average calculation time of 0.045 seconds for the 3D position of the stigma. This helps the robot arm to move the pollen brush to a close position on the stigma in a timely manner.

[0349] Servo pollination. To deliver pollen to the flower, the pollen brush must gently contact the pistil surface. The inventors achieved this using a coarse-to-fine servo strategy that combines a circular spiral servo strategy with a stigma contact confirmation model based on Inception-v3. Across 320 experiments, pollination success rates for k = 23 and k = 29 servo points were 94.45% and 99.68%, respectively (Figures 4H and 4I).

[0350] Finally, after successfully pollinating a flower, if there are still unprocessed flowers in the current image view, the system will initiate a new pollination process, processing these flowers through steps such as flower stigma perception, stigma pose estimation, and servo pollination. Once all mature flowers with exposed stigmas on a plant have been pollinated, the carrier will move the robot to the next plant. This pollination process will continue until all plants are pollinated.

[0351] Heterosis analysis

[0352] Mid-parent vigor is defined as the difference between the yield or mean value of a quantitative trait in a hybrid (F1) and the mean value (MP) of the same trait in both parents (P1 and P2), divided by the mean value of the same trait in both parents. Super-parent vigor is defined as the difference between the yield or mean value (F1) of a hybrid and the mean value (HP) of the same trait in the higher-valued parent, divided by the mean value of the higher-valued parent. Heterosis can be assessed by calculating mid-parent vigor and super-parent vigor. To assess heterosis in the first-generation hybrids of five hybrid combinations generated by GEAIR, the parental materials and first-generation hybrids were grown in a greenhouse and managed normally until harvest. Three traits were selected for statistical analysis: total yield of the first six fruit bunches, fruit weight, and Brix. The mean values ​​for each trait of the parents and hybrids were calculated based on the statistical data, and the mid-parent value was then calculated. Mid-parent heterosis was calculated according to the formula Mid-parent heterosis (%) = (F1-MP) / MP×100, and better-parent heterosis was calculated according to the formula Better-parent heterosis (%) = (F1-HP) / HP×100.

[0353] Determination of lycopene content by ultra-high performance liquid chromatography

[0354] Lycopene content was determined by ultra-performance liquid chromatography (UPLC) according to a previously reported method. 0.5 cm × 0.5 cm mesocarp sections of six red-ripe fruits from three plants were repeatedly cut, frozen in liquid nitrogen for at least 15 minutes, and stored in an ultra-low temperature freezer (−80°C). The samples were ground into powder using a tissue grinder (30 Hz, 60 seconds). Three replicates were collected from each group. Approximately 250 mg of fruit powder was weighed and mixed with 1 mL of a 2:1 mixture of chloroform and methanol. Then, 500 μL of a 50 mM Tris-HCl (pH 7.5) and 1 M NaCl solution was added, incubated on ice for 20 minutes, and centrifuged at 15,000 g for 10 minutes. The lower organic phase was collected, and the remaining supernatant was extracted again and concentrated using a nitrogen purge. The resulting residue was redissolved in 400 μL of methanol and subsequently analyzed. Lycopene was detected using an Agilent 1290 UPLC analytical platform. Chromatographic separation was performed using a YMC Carotenoid column (4.6 mm × 250 mm, 5 μm, Japan). The mobile phase used in this study consisted of methyl tert-butyl ether, methanol, and water (Solution A), and methanol alone (Solution B), with a gradient elution at a flow rate of 1 mL / min. UPLC peak areas were integrated at a wavelength of 450 nm and calibrated using a standard. The HPLC-grade lycopene standard (SMB00706-1MG) was purchased from Sigma-Aldrich.

[0355] Determination of flavor compounds in tomatoes using GC-MS

[0356] The content of major flavor compounds in tomatoes was determined by gas chromatography-mass spectrometry (GC-MS). For each sample, six ripe red fruits were collected from three plants, cut into small pieces, mixed thoroughly, and then quickly frozen in liquid nitrogen for at least 15 minutes. The frozen samples were stored in an ultra-low temperature freezer (-80°C). Before measurement, the samples were ground into powder using a tissue grinder (30 Hz, 60 sec). Each group included three technical replicates, and approximately 200 mg of fruit powder was weighed in the experiment. Volatile compounds were enriched by headspace solid-phase microextraction (HS-SPME) and then separated and detected by gas chromatography-mass spectrometry. 200 mg of tissue powder was mixed with 400 μL of 20% w / v NaCl solution and placed in a 4 mL glass vial (Agilent Technologies). The volatile compound 2-heptanone was added as an internal standard at a final concentration of 0.125 ng / μL. To collect volatiles, a 100 μm SPME fiber coated with DVB / CAR / PDMS (Sigma-Aldrich) was inserted into a preheated vial (30°C, 5 min) and exposed to the headspace at 30°C for 30 min. The SPME fiber was placed in the injector port of a GC-MS instrument (Agilent 7890A GC-5975C MSD) equipped with a DB-5MS column (30 m × 0.25 mm × 0.25 μm) and held for 30 sec. The oven temperature was initially maintained at 40°C for 1.5 min and then increased to 150°C at a rate of 5°C / min. The temperature was then increased to 260°C at a rate of 15°C / min and held for 10 min. The quadrupole mass spectrometer temperature was set to 150°C, and the mass spectrometer scan range was 50–400 m / z. Metabolites were tentatively identified by comparison of retention times and spectra with commercial standards or by spectral matching with entries in the NIST 08 library (NIST and Wiley libraries). Peak areas for each metabolite, obtained in SIM mode (signal-to-noise ratio greater than 5), were normalized to that of 2-heptanone before data processing. Relative amounts of target metabolites were calculated using 2-heptanone as an internal standard.

[0357] Data statistical analysis

[0358] Flowering time was determined by counting the number of leaves before the first inflorescence. Single fruit weight was measured using mature red fruits. Fruit set was calculated by dividing the number of fruits in the first two fruiting ears of a plant by the number of flowers in the first two inflorescences. Total fruit yield was the sum of the fruits in the first three ears of a plant, with the number of fruits at or after the color break stage accounting for at least 95% of the total fruit. Single fruit weight and yield per plant were determined using a Sartorius electronic balance. No fewer than 12 plants of each mutant were used for fruit set experiments in greenhouse and field cultivation. For analysis of stigma exsertion, the inventors manually measured the length of stamens and pistils from fully open flowers and calculated the pistil / stamen ratio. Stamen weakening and carpel number were analyzed using a Zeiss SteREO Discov V20@1126 microscope. The number of samples used for each genotype is indicated in the statistical graphs. All photographs were taken from representative individuals. For the above analyses, all data points were plotted as individual points in the boxplots. All data were analyzed using the mean ± standard deviation. The mean values ​​between groups were compared using a two-tailed two-sample t-test, and the p-values ​​were calculated using Microsoft Excel software.

[0359] Example 1: Tomato Flower Morphology and Design of the GEAIR Breeding System

[0360] A typical cultivated tomato flower has five to six alternating sepals and an equal number of yellow petals. The third whorl of floral organs is a cone of five or six yellow stamens, connected laterally by interwoven hairs. The cone is enclosed by two to three fused carpels, forming a multilocular ovary with an extended style and stigma (Figure 1A). Wild ancestors of tomato (such as Solanum pimpinellifolium) maintained genetic diversity through cross-pollination, with an exposed stigma not enclosed by other floral organs (Figures 1A and S1A). During crop domestication and improvement, artificial selection shifted reproductive strategies toward self-pollination, resulting in the production of nondisjunctive seeds that can maintain and fix desired traits, but at the expense of genetic diversity, leading to a genetic bottleneck (Figure 7A). Consequently, the initially exposed stigma gradually retracted (Figure 7A). However, because hybrid breeding relies entirely on manual detasseling and pollen smearing for cross-pollination, stigma-retracted germplasm incurs significant labor and time costs in hybrid breeding (Figure 7A). These labor and time costs account for a large portion of the hybrid seed production price.

[0361] In order to quickly produce male sterile lines with exposed stigmas and develop intelligent unmanned hybrid breeding strategies, the inventors proposed the GEAIR (Gene editing with artificial intelligence-based robots) breeding system (Figure 7B). The inventors used gene editing technology to target the B-class gene GLO2 in the ABC model of flower development. By changing the morphology of the stamens and exposing the stigmas, they produced male sterile lines with sterile pollen, thereby obtaining a flower morphological trait that is friendly to artificial intelligence robots. The inventors designed a robot system based on deep learning perception to achieve fully autonomous hybrid breeding in a greenhouse environment, which can eliminate manual labor and improve breeding efficiency.

[0362] Example 2: Obtaining a Tomato Male Sterile Line with Exposed Stigmas by Gene Editing GLO2

[0363] Some tomato B-class MADS-box gene mutants (such as sl2 and 7B-1) exhibit male sterility with exserted stigmas, but these mutants have suboptimal genetic backgrounds, and some also exhibit growth defects, making them unsuitable for modern, commercial, large-scale hybrid seed production. While RNAi can be used to reduce GLO2 gene expression, these transgenic lines exhibit exserted stigmas, but these are not completely male sterile and can produce deformed fruit due to factors such as carpelization of stamens. Therefore, the present inventors selected the superior cultivar Ailsa Craig and, by gene editing both the coding and non-coding sequences of its GLO2 gene, rapidly induced the exserted stigma male sterility trait (Figures 1B-D). Unlike RNAi lines, glo2-a1 fails to produce fertile pollen, exhibiting complete male sterility, while flowering time, plant height, and inflorescence structure remain unaffected (Figures 1E-H and 7E). However, only a small fraction of flowers exhibited exposed stigmas; even in flowers with exposed stigmas, the length of the exposed stigma was short, which was not suitable for the precise manipulation required by the robot, which requires the stigma to be strong and healthy and clearly exposed (Figure 1I and Figure 1J). Therefore, the loss-of-function glo2 allele was considered unsuitable.

[0364] In order to obtain a fine-tuned GLO2 allele in structural male sterility with minimal side effects and retain yield traits, the present inventors used a multiplex CRISPR-Cas9 gene editing system to design sgRNAs (Single-guide RNAs) targeting the second intron and 3'UTR (Figures 1C, D). The two mutants had large structural variations and were named glo2-dele and glo2-inver, respectively. The glo2-dele gene had a 3,225bp deletion from the second intron to the 3'UTR. glo2-inver produced a 2795bp inversion and a 420bp translocation between the second intron and the 3'UTR. These mutations altered the GLO2 open reading frame and may produce truncated proteins (Figures 1C and 7C). The other four mutants (designated glo2-in2-a1 to -a4) harbored different mutations at the target sites within the second intron, resulting in deletions of 598 bp (glo2-in2-a1) and 604 bp (glo2-in2-a3) (Figure 1D). In addition to a 1 bp insertion at the first target site, glo2-in2-a4 harbored deletions of 109 bp and 34 bp at the second and third target sites, respectively. glo2-in2-a2 exhibited a 104 bp deletion at the first target site and a 259 bp inversion at the second target site. None of these four mutations affected the protein coding region or open reading frame of GLO2 (Figures 1D and 7C). RT-qPCR revealed that GLO2 expression was downregulated in all mutants (Figure 7D). Compared to wild-type Ailsa Craig, expression of glo2-a1, glo2-dele, and glo2-inver was downregulated 10- to 30-fold (Figure 7D). glo2-in2-a2 (mutation only in the second intron, without destroying the coding region) is the most strongly downregulated mutant, with its transcriptional abundance nearly 100 times lower than that of Ailsa Craig (Figure 7D). The expression levels of glo2-in2-a1 and GLO2 were also downregulated in glo2-in2-a3 and glo2-in2-a4, but the downregulation was smaller (Figure 7D). Taking into account the regulatory interactions between class B genes, the inventors tested whether other class B genes were also downregulated. In different glo2 mutants, the other three class B MADS-box genes GLO1, DEF and TM6 were unaffected or downregulated to varying degrees (Figure 7F). The degree of downregulation of class B MADS-box genes was positively correlated with the degree of stamen defects observed in glo2 mutants (Figures 7D, E). The other three class B genes did not mutate, so this differential expression was not the result of Cas9 off-target.

[0365] In order to identify the ideal stigma-exposed male sterile line, the inventors screened various mutant phenotypes. The flowering time, plant height, inflorescence structure, and sepal and petal morphology of the glo2 mutant did not change, but the fertility of the stamens and pollen was affected (Figures 1E-H and 7E). Glo2-in2-a1 and glo2-in2-a3 developed normal stamens and produced fertile pollen like Ailsa Craig (Figures 1E and 7E). Glo2-in2-a4 showed normal stamens, fertile pollen, and indented stigma (Figures 1E and 7E). Glo2-a1, glo2-dele, glo2-inver, and glo2-in2-a2 flower organs produced curled stamens with no interwoven hairs between them, which could not connect to form a cone structure of the stamen group, did not produce fertile pollen, and showed complete male sterility (Figures 1E and 7E). Judging from the ratio of pistil length to stamen length, the stamens of these mutants were bent and shortened, exposing the stigma (Fig. 1I, J).

[0366] This complete male sterility and stigma exsertion phenotype was consistent under protected greenhouse and field cultivation conditions (Figures 1E, 7E and 8A, B). In summary, the present inventors rapidly generated a tomato male sterile line with exserted stigma by editing the non-protein coding regions of class B genes.

[0367] Example 3: Generation of a novel tomato male sterile line with exserted stigma without yield and seed quality loss

[0368] To evaluate whether the floral morphological changes of these glo2 mutants affect fruit yield and other breeding traits required for commercialization, the inventors used Ailsa Craig pollen to artificially pollinate three male sterile mutants with exposed stigmas (glo2-dele, glo2-inver, and glo2-in2-a2). The remaining three mutants that produce fertile pollen (glo2-in2-a1, glo2-in2-a3, and glo2-in2-a4) were self-pollinated and set fruit as controls. After artificial pollination, the fruit set rate of the three male sterile mutants was comparable to or higher than that of Ailsa Craig. Although the total fruit yield was not statistically significantly different from that of Ailsa Craig, except for glo2-in2-a2, the single fruit weight of glo2-dele and glo2-inver was higher than that of Ailsa Craig (Figures 2A-D). Although glo2-in2-a4 was self-fertile, its fruit weight was lower than that of Ailsa Craig (Figure 2C) and its fruit set rate was also reduced (Figure 2B). These data indicate that the female reproductive organs of the male-sterile mutants glo2-dele, glo2-inver, and glo2-in2-a2 are not affected and are suitable for hybrid breeding.

[0369] The carpelized phenotype resulting from stamen-to-carpel conversion often leads to deformed fruit, potentially impacting the commercial value of fresh tomatoes. Indeed, in some glo2 mutants, the converted carpelized stamens fuse with the central carpel, partially forming small, fruit-like carpels upon maturity, resulting in deformed fruit (Figure 2A). Based on the severity of stamen-to-carpel conversion, the present inventors categorized the mutant stamens into four categories: normal stamens (NO), stamens with carpelized structures on the front (CS), stamens with carpelized structures and externally attached ovules (EO), and stamens completely converted to carpels (TC) (Figures 2E and 8C). The degree of stamen carpelization in the three completely male-sterile mutants correlated with the rate of fruit deformity. Among the male-sterile mutants, glo2-in2-a2 exhibited the strongest carpelized phenotype, the highest TC percentage, and the highest fruit deformity rate (13.95%). glo2-dele and glo2-inver had relatively low fruit deformity rates, each around 2% (Figures 2A, F). The number of fruit ventricles was consistent with the number of carpels (Figures 2F, 8D, and 8E). The increased fruit size of the male sterile mutant after artificial pollination may be related to the increased number of ventricles (Figure 8D).

[0370] For commercial hybrid breeding, seed quality and fruit yield are equally important. Therefore, the present inventors evaluated the germination rate of hybrid seeds obtained from male sterile mutants after artificial pollination. The seed number, 100-seed weight and germination rate of glo2-dele and glo2-inver hybrids were no different from those harvested from Ailsa Craig, while the seeds of the glo2-in2-a2 hybrid germinated slightly slower (Figures 2G-2J). Among the three male sterile mutants, glo2-inver had a better stigma exposure trait and had the least impact on fruit yield and seed quality. Therefore, the present inventors selected glo2-inver as the object of further study.

[0371] Example 4: Flower Detection and Stigma Positioning Using Deep Learning

[0372] To evaluate the hybridization efficiency of male sterile lines with exposed stigmas, the inventors compared the time cost of hybridization with wild-type S. pimpinellifolium using the glo2-inver mutant and its background material, Ailsa Craig, as female parents. As shown in Figures 3A and 3B, the pollination time of hybridization using glo2-inver as the female parent was shortened by 54%, while fruit set did not differ significantly. Notably, the major obstacle to robotic manipulation of stigmas is that the stigma of typical domesticated tomatoes is hidden within a closed stamen cone. This was completely overcome by the gene-edited floral morphology, making robotic manipulation possible. Leveraging this carefully designed floral morphology, the inventors further developed an autonomous robotic pollination system for the production of F1 hybrids of tomato.

[0373] The biggest challenge in automated pollination lies in detecting and locating the stigma of the target flower, as it can be obscured by leaves, branches, fruit, and other flowers. Stigmas can also vary in orientation, and the precise angle of orientation determines the approach trajectory that the robotic pollen brush must follow to achieve pollination. Relying on specialized knowledge to design such a flower-stigma estimation algorithm is difficult; instead, the inventors adopted a data-driven approach, training a deep neural network to detect flowers requiring pollination. Specifically, the inventors collected 12,000 images, annotating the bounding boxes, segmentation masks, and orientations of all flowers requiring pollination in each image. The inventors used 9,600 images for deep neural network training, and the remaining 2,400 images for evaluating the trained model. Based on the "You Only Look At CoefficienTs" (YOLACT) deep learning model, the inventors added a network branch for object detection orientation to infer each flower's bounding box, segmentation mask, and pistil orientation from camera images, denoted here as YOLACT_Orient (Figures 3C, 9A, and 9B). The inventors benchmarked YOLACT_Orient and other state-of-the-art detection methods, such as Mask Region-based Convolutional Network method (Mask R-CNN) and DEtection Transformer (DETR), on 128 male-sterile tomatoes with exposed stigmas (Figures 9A, B). YOLACT_Orient achieved the highest flower detection accuracy of 82% and the fastest inference time of 60ms per frame (Figure 3D). The inventors evaluated it on a low-end computer equipped with a 2.60GHz central processing unit (CPU), 8.0GB random access memory (RAM) and an NVIDIA RTX 2080Ti 16GB graphics processing unit (GPU). It should be noted that the inventors only marked flowers that could be pollinated and ignored immature flowers without stigmas. Therefore, after training, the deep neural network can distinguish between flowers that are suitable for pollination and those that are not suitable for pollination. After detecting a flower, the pollination robot delivers pollen to the target stigma in two steps: 1) aligning the end effector holding the pollen and placing it parallel to the direction of the flower's stigma; and 2) moving the pollen brush and gently pushing it toward the flower until it touches the stigma.The BrambleBee pollination robot is trained using artificial flowers, and the direction of the flowers is divided into only three categories: left, front, and right (Strader, J., et al., Flower Interaction Subsystem for a Precision Pollination Robot. In 2019 IEEE / RSJ International Conference on Intelligent Robots and Systems (IROS) 5534-5541. DOI: 10.1109 / IROS40897.2019.8967752). Here, the inventors divide the pistil direction into five categories: left, right, front, upward, and downward to better reflect the diversity of the flower pose and obstacles in the scene (Figure 9A). The inventors manually scored the direction of the flower images and evaluated the YOLACT_Orient model for flower direction estimation (Figures 3C and 9A). The detection accuracy rates for the front, left, right, top, and bottom directions of the pistil were 89.6%, 81.9%, 79.6%, 93.6%, and 83.9%, respectively, and the recall rates were 90.7%, 80.6%, 77.9%, 96.4%, and 82.8%, respectively, all of which were better than BrambleBee (Figure 3E).

[0374] Precise pollination requires determining the three-dimensional position of the stigma. On the one hand, the stigma is very small, occupying only a small portion of the captured image. On the other hand, pollination requires the pollination brush to be accurately and gently placed on the stigma surface. This places stringent demands on the accuracy of estimating the stigma's three-dimensional position. To meet this requirement, the present inventors designed a pseudo-binocular ranging strategy and a feature-based position estimator (Figure 3F). First, our system captures two images facing the plant at two locations parallel to the plant, 10 cm apart. Second, the 3D position of the stigma is determined using feature point matching using Speeded Up Robust Features (SURF) and Random Sample Consensus (RANSAC). The present inventors evaluated our algorithm based on real stigma positions measured on 64 randomly selected tomato flowers. The average localization errors of the stigma along the x, y, and z dimensions were 5.0 mm (maximum distance 13.3 mm), 8.5 mm (maximum distance 22.2 mm), and 5.3 mm (maximum distance 12.9 mm), respectively (Figure 3G). The present inventors processed only the portion of the image obtained by 2D detection that was cropped to the petal region. As a result, feature matching and RANSAC were highly robust to background noise and computationally fast (the average time to compute the 3D position of the stigma was 0.045 s) (Figure 3G), enabling the robot to move the pollen brush to the approximate position of the stigma with almost no delay.

[0375] Example 5: Implementing Robotic Automatic Pollination of Tomato Plants

[0376] There are two significant challenges in achieving robotic pollination. First, the pollination brush needs to approach all mature flowers on the plant from multiple directions. Second, stigmas are small and fragile, and can only withstand slight forces for a reasonable period of time. The inventors took these factors into consideration when designing the pollination mechanism, as detailed below.

[0377] The robot consists of a mobile base (Carrier / ShiHe MR1000 vehicle), an ultra-wideband positioning module, a RealSense D435i camera, a UR5 robotic arm (pollination arm), a Robotiq F85 gripper (pollination gripper), a soft brush (pollination brush) and a pollen container (Figures 4A-4C). The robot base safely moves and turns between rows of tomato plants, with each row 2m apart (Figure 4D). Using the positioning function of LocalSense and the map-based scanning and positioning (STL) navigation strategy, the robot can locate and navigate to the specified target tomato plant within a range of 10cm (Figures 4E-F). In addition, the inventors have also designed the robot base and robotic arm so that the pollination brush can reach the stigma that needs to be pollinated when facing the tomato plant in the three-dimensional operating space of the robotic arm.

[0378] After the pollination robot arrives at the tomato plant, YOLACT_Orient and pseudo-binocular ranging (SURF+RANSAC) methods will obtain the initial three-dimensional position of the stigma from the captured image. The pollination work involves accurately guiding the pollination brush to the stigma surface after determining its initial three-dimensional position. The inventors adopted a ring spiral servo strategy to resist the detection error of the stigma position estimator, with an error range of 4-25mm (Figure 4G). This range can cover the three-dimensional stigma position prediction error in the SURF+RANSAC step (maximum ~22.2mm). This capability reflects the high-precision sensory-motor-control skills of human operators in such operation tasks. In addition, the inventors also used a lightweight neural network Inception-v3 to capture and classify the stigma in real time through camera images, that is, whether the stigma is currently contacted by the pollen brush, thereby ensuring that true contact is achieved between each pistil and the pollen brush (Figures 4H and 9C-F). Within an acceptable time frame, the pollination brush can move through up to 36 servo points. At the same time, if stigma contact is confirmed based on Inception-v3, the system will stop servoing the current stigma. In 320 experiments, the pollination success rates of k = 23 and k = 29 servo points reached 94.45% and 99.68%, respectively (Figure 4H, Figure 4I), exceeding the 76.9% pollination success rate of BrambleBee when using artificial flowers.

[0379] In the present inventors' experiments, tomato plants were evenly arranged at known positions within each row, and four LocalSense gateways, 2.9 meters above the ground, were installed at the four corners of a 12 m × 9.8 m greenhouse. The robot carried LocalSense tags that continuously estimated its position and orientation in real time. The robot autonomously navigated and pollinated within the greenhouse, achieving a fruit set rate of 74.6% ± 10.7% (Figure 4J). Manual pollination by a human operator achieved a fruit set rate of 85.3% ± 22.2%, which was not significantly different from the robot's autonomous pollination. Notably, the robot's autonomous pollination achieved a more stable fruit set rate, as it was unaffected by differences in professional training between expert and novice pollinators (Figure 4J), as evidenced by its smaller variance. In summary, the pollination robot navigated to a plant, detected a sufficiently mature unpollinated flower, estimated the position and orientation of the stigma, guided the pollination brush to pollinate, and repeated this task for each flower. The robot autonomously repeated this process for each plant (Figures 4K and 9G).

[0380] In order to apply GEAIR to the production of F1 hybrid seeds, the inventors used the above-mentioned glo2-inver (Alisa Craig background) as the female parent and five different inbred elite varieties as the male parents, including M82 (processing tomato variety), Beijing1 (local traditional variety with excellent flavor) and three varieties with considerable stress resistance and good flavor (TS545, TS181 and TS590). F1 hybrids were obtained by GEAIR robot pollination. Compared with artificial hybridization using wild-type Alisa Craig as the female parent, GEAIR greatly saves labor costs and time (Figure 4L). The F1 offspring of the four male parents hybridized with glo2-inver showed significant population consistency and hybrid vigor, and the fruit yield and quality (Brix, fruit sugar content) were improved (Figure 10).

[0381] Example 6: Cultivation of stress-resistant tomatoes using GEAIR by combining de novo domestication with rapid breeding

[0382] Desirable traits such as stress tolerance, ideal flavor, and high nutritional value are complex, making breeding tomato varieties that combine these traits challenging and time-consuming. Inbreeding of high-quality cultivated varieties can reduce genetic diversity, creating genetic bottlenecks and further complicating the development of superior tomato varieties. Wild tomatoes possess excellent stress tolerance, flavor, and nutritional properties, but introducing these traits into high-quality cultivated varieties is time-consuming and labor-intensive. De novo domestication of wild species has been shown to be an alternative strategy, but the limitations of gene editing tools, particularly the low efficiency of plant gene knock-in technology, make rapid domestication of some yield traits controlled by complex loci difficult. Therefore, a compromise approach is to domesticate wild species with desirable traits de novo, making them as similar as possible to high-quality cultivated varieties, and then hybridize them with high-quality cultivated varieties to introduce the desired traits. This approach can greatly improve the efficiency of introducing stress tolerance traits and selecting for desirable traits in offspring.

[0383] Rapid breeding utilizes LED lighting and a 22-hour day length to optimize photosynthesis and promote rapid growth. Six generations of wheat can be produced annually, compared to two using traditional breeding methods. However, insects' vision and navigation systems fail under LED lighting, making it difficult to find flowers and pollinate as they do naturally. Artificial pollination is time-consuming and costly. These factors hinder the application of hybrid breeding in rapid breeding programs.

[0384] To combine the advantages of de novo domestication and rapid breeding, the present inventors combined GEAIR with de novo domestication in an LED-based artificial climate chamber to rapidly cultivate new tomato germplasm with strong stress resistance, rich flavor, and abundant nutrition (Figure 5A). The present inventors selected S. pimpinellifolium as the starting material because it is tolerant to saline and alkali, and resistant to major tomato diseases such as powdery mildew and bacterial scab, while also exhibiting rich flavor, high lycopene content, and a high Brix value. Using de novo domestication, the present inventors generated germplasm with photoperiod insensitivity and compact plant structure as the male parent line (a fundamental characteristic of modern tomato cultivars). The aforementioned glo2-inver male sterile line with exposed stigmas based on the Ailsa Craig background was used as the female parent line. The present inventors used GEAIR for hybridization and cultivated the progeny in an LED growth chamber, allowing tomato plants to undergo four generations within a year, significantly reducing the iteration time (Figure 5A). The present inventors first screened individuals with desirable traits from the F2 population (Figures 11A-F). For example, the powdery mildew infection rate of the F2 population was 36.9%, while the infection rate of the Ailsa Craig parent was 94.3%, indicating that the hybridization was successful (Figure 11G). Within the F2 population, the inventors selected plants that were resistant to powdery mildew and exhibited a bideterministic plant structure. From the F3 population, the inventors identified plants that were tolerant to saline-alkali soil (salinity of 0.3% (w / v), pH 8.0) (Figures 11H-I). These selected plants were then planted to generate an F4 population, from which lines with excellent traits and high population uniformity were obtained for further testing of fruit yield, quality, and flavor. After four rapid generations, the inventors identified an F4 inbred line with excellent traits and relative stability (Figures 5B-C). These plants were grown in a greenhouse for trait analysis, and their phenotypic consistency was high, with fruit yield comparable to that of the Ailsa Craig maternal line but higher than that of the S. pimpinellifolium parent (Figure 5D). The fruits of this line were larger than those of S. pimpinellifolium but smaller than those of Ailsa Craig (Figure 5E). Remarkably, the Brix content of the fruits increased by 61.85%, and the lycopene content increased by 50.7% (Figures 5F,G), indicating that it had inherited the desirable traits of the parent line. Notably, the contents of key flavor-related volatiles were also increased in these fruits compared to Ailsa Craig (Figures 5H-5P). For example, 2-isobutylthiazole, a major component of tomato aroma, was ninefold higher than in Ailsa Craig, suggesting that flavor lost during domestication has been restored (Figure 5N).These results suggest that combining GEAIR with de novo acclimation and using climate chambers equipped with full-spectrum LEDs saves time and labor while facilitating the rapid development of stress-resistant varieties with unique flavors that meet the tomato needs of growers and consumers.

[0385] Example 7: Suitability and potential of GEAIR for modifying flower morphology in other Solanaceae crops

[0386] To assess the broad applicability of GEAIR to other Solanaceae crops with similar floral structures (including peppers, eggplants, and potatoes), the present inventors analyzed the floral morphology of representative species of the Solanaceae family ( FIG6A ). Flower morphology can generally be divided into three groups: stamen fusion (tomato, wild tomato), intermediate morphology (eggplant, potato), and stamen separation (pepper, mushroom, petunia, tobacco) ( FIG6B ). Despite the morphological differences, a gradual transition from stamen fusion to stamen separation can be observed, with stamen fusion being a more recent trait in evolution, while the stigma retraction trait is more recent.

[0387] Given the role of GLO2 in stamen fate determination, the inventors selected 24 Solanaceae species, including 7 species with fused stamens, 11 species with intermediate morphology, and 6 species with detached stamens, and reconstructed a phylogenetic tree based on GLO2 protein sequences (Figure 6C). The differentiation of GLO2 largely reflects the diversity of stamen morphology (Figures 6C and 12A). GLO2 is divided into two groups, Group A representing fused stamens and Group B representing detached stamens. Intermediate morphological types belong to either group, suggesting that other class B MADS proteins may be involved (Figures 6C and 12A). The inventors then analyzed the differentiation of three other class B MADS-box proteins, GLO1, DEF, and TM6, and found that DEF and TM6 showed characteristics similar to GLO2 (Figures 12B-D), also reflecting the differentiation of stamen types. These results are consistent with previous studies on the variability of floral homologous class B genes, namely that mutations in the keratin domain lead to the formation of different complexes between class B proteins, which may explain the differentiation of stamen morphology in Solanaceae plants. The overall conservation of the multimeric class B MADS-box protein complex provides potential target genes, such as GLO2, DEF, and TM6, for flower morphology design to achieve robotic pollination.

[0388] Example 8: Creation of a soybean male sterile line with exposed stigma

[0389] Cultivated soybean is a typical self-pollinating crop with closed flowers. Soybean flowers are highly closed, with the stigma tightly encased by the keel, making it difficult to accept foreign pollen and achieve outcrossing (Figure 13A). This characteristic makes soybean hybridization extremely time-consuming and labor-intensive, and each hybridization typically yields only one or two seeds. Due to the high cost of artificially producing soybean hybrid seeds, the utilization of hybrid vigor in soybean has been stagnant. To overcome this dilemma, it is necessary to modify soybean floral morphology and create male-sterile lines with exserted stigmas. Class B MADS-box genes involved in floral development regulate the development of petals and stamens in flowering plants, and this may also be conserved in soybean. Editing class B MADS-box genes in soybean is expected to produce male-sterile lines with exserted stigmas (Figure 13B). Soybean is a paleotetraploid plant with a high number of double-copy genes in its genome, and class B MADS-box genes are no exception. By homologous comparison, it was determined that there were 8 B-class MADS-box genes in soybean, including GmPI1, GmPI2, GmPI3, GmPI4, GmAP3a, GmAP3b, GmTM6a and GmTM6b. The inventors designed targets for 6 of these genes and performed gene editing, and obtained some mutant materials (Figure 13C). The inventors carefully observed the flower organ phenotypes of the mutant materials in the T1 generation, including flowers, sepals, petals, stamens and pistils of gmpi1 gmpi3+ / -, gmpi3, gmpi1 gmpi3, gmap3b gmtm6b and gmap3a gmap3b gmtm6b mutants (Figure 13D-I). The inventors found that the flower organ structure of gmpi1 gmpi3+ / -, gmpi3, and gmpi1 gmpi3 mutants did not differ significantly from that of the wild type (Figure 13D-G). In contrast, in the gmap3b gmtm6b mutant, 25% of flowers had detached keels, sometimes with one additional petal, resulting in exserted stamens and pistils (Figure 13H), but fertility remained normal. Surprisingly, the inventors observed the desired phenotype in the gmap3a gmap3b gmtm6b mutant, with sepalized petals, carpelized stamens, and exserted stigmas (Figure 13I). Specifically, the petals of the gmap3a gmap3b gmtm6b mutant were pubescent on the abaxial surface, yellow-green in color, shorter than the sepals, and failed to elongate in the later stages. The keels were detached, and the stamens exhibited a strongly carpelized phenotype, with pubescent ovules and stigma-like structures visible in severe cases. However, the calyx and pistil of the mutant appeared normal (Figure 13I). The leaves and stems of the gmap3a gmap3b gmtm6b mutant were identical to those of the wild type (Figures 13J-K). Under normal cultivation conditions, the mutant could not produce pods like the wild type, indicating that it was completely male sterile (Figure 13K). After artificially assisted cross-pollination, the mutant could produce pods normally, indicating that its pistil fertility was normal (Figure 13L).The above results show that by editing the B-type MADS-box gene, male sterile soybean varieties with exposed stigmas can be quickly created, eliminating flower morphological obstacles, facilitating hybridization, and making automated soybean hybrid seed production and commercial hybrid advantage utilization possible.

[0390] discuss

[0391] In agriculture, artificial intelligence and robotics are already being used for environmental data collection, phenotyping, fruit harvesting, weeding, pest and disease monitoring and control, and fertilizer application. These applications rely primarily on optimizing robot design to meet the needs of different crops and cultivation methods. However, the use of robots for precise crop manipulation is relatively limited. A major reason for this situation is that crop traits have not yet been perfectly aligned with intelligent industrialization. For example, to match intelligent industrialization, there has been a shift from optimizing for high yield per plant to achieving more compact crops with highly synchronized flowering and ripening times. This shift is driven by the agricultural industry's focus on increasing yield per unit area and achieving dense planting to facilitate mechanical management and harvesting.

[0392] Some agronomic traits derived from the development of domestication and breeding are a major obstacle to the application of artificial intelligence and robots. For example, the structure of floral organs determines the pollination and fertilization characteristics of crops (Figure 6A, B). To ensure the genetic consistency of seeds, the exposed stigmas of wild plants are domesticated and artificially selected and enclosed by stamens or other floral organs. In addition, manual emasculation makes the retracted stigmas and female reproductive organs easily damaged, resulting in sterility. If the morphology of flowers is not properly adjusted, it will be difficult for intelligent robots to carry out hybrid breeding, and manual labor will still be required.

[0393] In order to achieve robot-led automated pollination, the floral organ structure needs to be modified, for example, in domesticated tomatoes, the stigma is exposed. However, floral organ structure often co-evolves with other reproductive traits, so specific changes are required without affecting these other traits. Historically, breeding materials with altered floral organ structure are usually obtained through natural variation or chemical or physical mutagenesis, followed by multiple rounds of genetic screening. This process is extremely time-consuming, and introducing these traits into new breeding materials still requires lengthy hybridization and separation to purify the genetic background. If a suitable target gene can be found, the rise of gene editing technology provides a realistic means to quickly produce ideal traits directly in a suitable genetic background.

[0394] Designing customized floral organ structures using the ABC model for breeding purposes has been a long-standing goal. Loss-of-function mutants of ABC model genes exist in various plant species. However, some mutants exhibit abnormal floral organs and suffer varying degrees of impairment in yield or fruit appearance. Consequently, the application of these genes in breeding has been hindered. This study aimed to rapidly create robot-friendly traits and enable automated pollination in tomatoes. Hybrid breeding in this species is labor-intensive and time-consuming due to manual detasseling and pollination. The inventors selected the class B gene GLO2 as a target gene because it controls stamen specialization in tomato. Using a multiplex gene editing system, they edited the non-protein-coding region to create a desired loss-of-function mutant with an exposed stigma. This method generated male sterile lines with altered stamen morphology and exposed stigma, significantly reducing labor input in hybrid breeding and overcoming the primary obstacle to the application of artificial intelligence and robotics in hybrid tomato breeding. To this end, the inventors developed an intelligent robot based on deep learning and algorithm optimization. The robot can capture and rapidly cross-pollinate male-sterile flowers in real time, with a success rate comparable to that of skilled human pollinators. It was trained using real tomato flowers in a realistic greenhouse environment that simulates a commercial setting. The robot saves manpower, improves breeding efficiency, and reduces breeding costs. The inventors demonstrated the versatility of the system by integrating GEAIR with de novo domestication and combining it with an optimized artificial climate growth space, thereby greatly accelerating tomato breeding using wild relatives as a source of genetic diversity. This strategy provides a comprehensive and realistic solution for developing new crops that are resilient to climate change, high quality, flavorful, and nutritious. The application of artificial intelligence-robots in breeding will safeguard agricultural efficiency and sustainable development to ensure future global food security.

[0395] This study provides an example for AI-assisted breeding. The inventors combined gene editing and artificial intelligence technology to innovate the plant-machine "interface" and realize automated pollination in tomato hybrid breeding. In addition to plants belonging to the GLO2 A class in the Solanaceae family, this strategy can also be extended to other major crops. For example, the stigma of cultivated soybeans (closed-flowered self-pollination) is tightly wrapped by the keel petals, making it difficult to accept foreign pollen. Since artificial emasculation and cross-pollination require huge labor costs, this flower morphology hinders soybean hybrid breeding. ABC-type gene editing can make the stigma exposed by changing the characteristics of the flower organs, simulating the automated pollination of the tomato male sterile line with exposed stigma in this study.

[0396] Partial sequence information:

Claims

1. A method for producing a modified plant, said method comprising targeted modification of at least one endogenous B-class gene of the ABC model of floral development of said plant, said modification resulting in a stigma-exserting phenotype in said modified plant.

2. The method of claim 1, wherein the plant is a Solanaceae or Fabaceae plant.

3. The method of claim 2, wherein the Solanaceae plant is selected from the group consisting of Solanum lycopersicum (tomato), Nicotiana benthamiana (tobacco), Capsicum annuum (pepper), Physalis pruinosa (mushroom), Solanum melongena (eggplant), Solanum tuberosum (potato), Solanum pennellii, Solanum chilense, Solanum habrochaites, Solanum pimpinellifolium, Solanum galapagense and Petunia hybrid (petunia), preferably, the Solanaceae plant is tomato, more preferably, the Solanaceae plant is selected from the group consisting of tomato cultivars Ailsa Craig, M82, Beijing1, TS545, TS181, TS590, most preferably, the plant is tomato cultivar Ailsa Craig; or the legume plant is selected from the group consisting of soybean (Glycine max), peanut, bean, pea, adzuki bean, mung bean, cowpea, kidney bean and lentil, preferably, the legume plant is soybean.

4. The method of any one of claims 1-3, wherein the at least one class B gene is GLO2.

5. The method of claim 4, wherein i) the GLO2 gene encodes a GLO2 protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence of SEQ ID NO: 1; ii) the GLO2 gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 2; and / or iii) the GLO2 gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence shown in SEQ ID NO:

3.

6. The method of any one of claims 1 to 5, wherein the at least one Class B gene is selected from soybean GmPI1, GmPI2, GmPI3, GmPI4, GmAP3a, GmAP3b, GmTM6a and GmTM6b genes, or a combination thereof; preferably, the at least one Class B gene is selected from soybean GmAP3a, GmAP3b, and GmTM6b genes, or a combination thereof; more preferably, the at least one Class B gene includes soybean GmAP3a, GmAP3b, and GmTM6b genes.

7. The method of claim 6, wherein i) the GmAP3a gene encodes a GmAP3a protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence of SEQ ID NO: 29; and / or the GmAP3a gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 30; and / or the GmAP3a gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 31; ii) the GmAP3b gene encodes a GmAP3b protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence of SEQ ID NO: 33; and / or the GmAP3b gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 34; and / or the GmAP3b gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence of SEQ ID NO: 35; and / or iii) the GmTM6b gene encodes a GmTM6b protein having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the amino acid sequence shown in SEQ ID NO:41; and / or the GmTM6b gene comprises a coding sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence shown in SEQ ID NO:42; and / or the GmTM6b gene comprises a genomic sequence having at least 75%, at least 80%, at least 85%, at least 90%, at least 95%, at least 96%, at least 97%, at least 98%, or at least 99% sequence identity to the nucleotide sequence shown in SEQ ID NO:

43.

8. The method according to any one of claims 1 to 7, wherein the modification is a substitution, deletion and / or addition of one or more nucleotides; or, the modification is a rearrangement such as an inversion of a fragment of the gene.

9. The method of any one of claims 1 to 8, wherein the modification is performed by introducing into the plant a gene editing system that targets the at least one class B gene.

10. The method of claim 9, wherein the gene editing system is a CRISPR, ZFN or TALEN-based gene editing system; preferably, the gene editing system is a CRISPR-based gene editing system.

11. The method of claim 9, wherein the method for introducing the gene editing system into the plant is selected from the group consisting of: gene gun method, PEG-mediated protoplast transformation, Agrobacterium-mediated transformation, plant virus-mediated transformation, pollen tube channel method, and ovary injection method.

12. The method of any one of claims 8-11, wherein the gene editing system targets a non-coding region of the GLO2 gene.

13. The method of claim 12, wherein the gene editing system targets the second intron and / or 3' UTR (3' untranslated region) of the GLO2 gene.

14. The method of any one of claims 12-13, wherein the gene editing system comprises a target sequence selected from one of SEQ ID NOs: 4-11 or a combination thereof.

15. The method of any one of claims 1 to 14, wherein the modification results in a mutated GLO2 gene of one of SEQ ID NOs: 12 to 18.

16. The method of any one of claims 1 to 15, wherein 1) the modification results in the plant expressing a truncated GLO2 protein; 2) the modification results in downregulation of expression of the GLO2 gene in the plant by at least 2-fold, at least 5-fold, at least 10-fold, at least 20-fold, at least 30-fold, at least 40-fold, at least 50-fold, at least 100-fold or more; 3) the modification does not affect pistil development in the modified plant; and / or 4) The modification results in impaired stamen development and the failure to produce fertile pollen in the modified plant.

17. The method of any one of claims 9-11, wherein the gene editing system targets the coding region of GmPI1, GmPI2, GmPI3, GmPI4, GmAP3a, GmAP3b, GmTM6a, GmTM6b or any combination thereof, preferably the coding sequence at the 5' end.

18. The method of claim 17, wherein the gene editing system comprises a target sequence selected from one of SEQ ID NOs: 22-23, 27-28, 32, 36, 40, and 44, or a combination thereof.

19. The method of any one of claims 1-11 and 17-18, wherein the modification results in a mutated class B gene of one of SEQ ID NOs: 45-56, or any combination thereof.

20. The method of any one of claims 1-11 and 17-19, wherein the modification results in mutated GmAP3a, GmAP3b and GmTM6b genes, preferably, the modification results in a mutated GmAP3a gene as shown in SEQ ID NO: 49 or 50, a mutated GmAP3b gene as shown in SEQ ID NO: 51 or 52, and a mutated GmTM6b gene as shown in SEQ ID NO: 55 or 56.

21. The method of any one of claims 1 to 20, wherein the modification results in a male sterile phenotype in the modified plant, preferably, the modified plant has a completely male sterile phenotype, more preferably, the modified plant does not produce fertile pollen.

22. The method of any one of claims 1 to 21, wherein the modified plant has comparable flowering time, plant height and / or inflorescence architecture compared to an unmodified wild-type plant.

23. The method of any one of claims 1-22, wherein after pollination, the modified plant has comparable fruit quality and / or fruit yield compared to an unmodified wild-type plant; for example, the modified plant has a proportion of deformed fruit of less than about 10%, less than about 5%, less than about 4%, less than about 3%, less than about 2%, less than about 1% or even less.

24. claim claim 1-23 among the method for any one, wherein said modified plant has equal ability of accepting pollen and / or column head accepts pollen ability unaffected compared with unmodified wild-type plant; For example, the column head of said modified plant can normally produce hybrid fruit after accepting pollen, and its fruit setting rate is suitable with wild type.

25. The method of any one of claims 1 to 24, wherein after pollination, the modified plant has comparable seed number, 100-seed weight, and / or seed germination rate as compared to an unmodified wild-type plant.

26. A modified plant produced by the method of any one of claims 1 to 25.

27. Use of the modified plant according to claim 26 in hybrid breeding, preferably, the modified plant is used in automated hybrid breeding.

28. A method for automated plant hybrid breeding using an automatic pollination robot, the method comprising: i) providing a plant having a flower with an exserted stigma phenotype as a hybrid mother plant; ii) enabling the automatic pollination robot to approach the plant and obtain an image of the plant; iii) enabling an automatic pollination robot to process the image using an artificial intelligence method, detect the flowers of the plant, and determine the three-dimensional position of the stigma; and iv) causing an automatic pollination robot to pollinate the stigmas in the flowers of the hybrid female plant with pollen from the hybrid male plant.

29. The method of claim 28, wherein the method further comprises v) harvesting hybrid seeds from the hybrid female plant.

30. The method of claim 28, wherein the hybrid mother plant in step i) is the modified plant of claim 24.

31. The method of any one of claims 28 to 30, wherein the automatic pollination robot comprises a movable base; a robotic arm assembly mounted on the base; a pollination device, a camera, and a pollen box, each detachably attached to the robotic arm assembly; Positioning module; and controllers.

32. The method of claim 31, wherein in step ii), the controller locates and navigates the movable base to the target hybrid female plant to be pollinated through the positioning module, for example, through the positioning function of LocalSense and a map-based scanning positioning strategy.

33. The method of claim 31, wherein in step iii), the automatic pollination robot processes the image to detect the flowers of the plant and determine the three-dimensional position of the stigma by: A) Input image containing flowers; B) Send the input image to the backbone network for feature extraction to obtain a feature map; C) sending the obtained feature map to the detection branch to obtain prediction coefficients, and sending the obtained feature map to the segmentation branch to obtain a prototype mask, wherein the obtained prediction coefficients include coefficients representing the flower orientation; and D) The obtained prototype mask is combined with the prediction coefficients, and the output result is obtained together with the coefficient representing the flower orientation through the cropping and threshold modules.

34. The method of claim 33, wherein in step A), the image is preprocessed, wherein the size of the input image is adjusted to W×H×3, wherein the pixel resolution ratio of W and H is set to 550.

35. The method of claim 33 or 34, wherein step B) comprises: b1) performing feature extraction on the input image to obtain a feature backbone consisting of feature maps c1 to c5; and b2) A feature pyramid network (FPN) consisting of feature maps p3 to p7 is obtained from the obtained feature maps c1 to c5.

36. The method of claim 35, wherein in step b1), five feature maps c1 to c5 with sizes from large to small are generated by a fully convolutional network (FCN).

37. The method of claim 35 or 36, wherein in step b2), the feature map c5 with the smallest size is subjected to a convolution layer to obtain the feature map p5; the feature map p5 is amplified by a bilinear interpolation and added to the convolved feature map c4 to obtain the feature map p4; the feature map p4 is amplified by a bilinear interpolation and added to the convolved feature map c3 to obtain the feature map p3; the feature map p5 is convolved to obtain the feature map p6; the feature map p6 is convolved to obtain the feature map p7.

38. The method of any one of claims 35 to 37, wherein step C) comprises: c1) Send all feature maps p3 to p7 of the Feature Pyramid Network (FPN) to the detection branch to obtain the coefficient c+b+o for each anchor box, where the detection branch includes a prediction head and a non-maximum suppression module (NMS). The coefficient c represents the classification confidence, the coefficient b represents the bounding box regression, and the coefficient o represents the flower orientation. c2) At the same time as step c1), the maximum feature map p3 of the feature pyramid network (FPN) is sent to the segmentation branch parallel to the detection branch, where m prototype masks are obtained.

39. The method of any one of claims 33 to 38, wherein the flower orientation refers to the direction of the flower stigma, including the following five categories: left, right, front, top and bottom.

40. The method of any one of claims 33 to 39, wherein the loss function is represented by the following formula (1): L = w cls ·L cls +w box ·L box +w mask ·L mask +w orient ·L orient (1) in, L cls Represents the classification confidence loss of whether the target category is a flower, and its weight coefficient is w cls , L box Represents the bounding box regression loss, whose weight coefficient is w box , L mask Represents the mask loss, and its weight coefficient is w mask , L orient Represents the flower orientation loss, and its weight coefficient is w orient .

41. The method of claim 40, wherein the flower orientation loss L orient It is represented by the following formula (2): in, θ represents the direction angle of the flower stigma frame. For example, the value of θ is set to 42. The method of any one of claims 33-41, wherein the output result obtained in step D) is (x, y, w, h, c, m, o), where x, y, w, h represent the position, width, and height of the flower box, c represents the flower classification confidence, m represents the number of prototype masks, and o represents the flower's top, bottom, left, right, front, and other directions.

43. The method of any one of claims 28 to 42, wherein step iii) further comprises using a "pseudo-binocular" method to perform three-dimensional positioning of the stigma.

44. The method of claim 43, wherein images of the plant stigma are acquired at two different positions, and the three-dimensional positioning of the crop stigma is acquired based on matching of feature points between the images.

45. The method of any one of claims 28 to 44, wherein in step iv), the controller drives the pollination device and the pollen box carried by the robotic arm assembly to move in front of the stigma based on the three-dimensional positioning of the stigma.

46. ​​The method of any one of claims 28 to 45, wherein in step iv), after the pollination device moves to the front near the stigma, the controller drives the pollination device carrying the pollen to move sequentially between a plurality of points in a set three-dimensional space including the three-dimensional positioning of the stigma until the pollination device coats the stigma with pollen.

47. The method of claim 46, wherein the three-dimensional space is set as a stereoscopic space that covers a maximum positioning error of the three-dimensional positioning of the crop stigma obtained by matching the feature points in a three-dimensional coordinate system.

48. The method of any one of claims 46 to 47, wherein the pollination device is first moved to the three-dimensional position of the stigma obtained by matching the feature points, and then gradually moves from near to far through the multiple points to the outer contour of the three-dimensional space.

49. The method of claim 48, wherein the three-dimensional space is configured as a cylinder, the center of mass of the cylinder is the three-dimensional position of the crop stigma obtained by matching the feature points, and the pollination device is driven to move sequentially along a plurality of points on an annular spiral line in at least two circular cross-sections of the cylinder until the pollination device coats the crop stigma with pollen.

50. The method of claim 49, wherein the span between adjacent said annular spirals is determined by said maximum positioning error.

51. The method according to any one of claims 46 to 50, wherein an image including the stigma and the pollination device is acquired each time the pollination device moves to one of the multiple points, and whether the stigma and the pollination device are in contact is determined based on their relative positions in the image, and when it is determined twice in succession that the crop stigma and the pollination device are in contact, it is determined that the pollination device has coated pollen on the stigma.

52. The method of claim 51, wherein a lightweight network architecture based on Inception-v3 is used to confirm whether the stigma is in contact with the pollination device.

53. The method of any one of claims 28-52, wherein the male hybrid plant is a plant that can produce fertile pollen and can pollinate the female hybrid plant; preferably, the male hybrid plant is a plant with excellent agronomic traits.

54. The method of any one of claims 28-53, wherein the hybrid father plant is Solanum lycopersicum (tomato), for example, the hybrid father plant is selected from the group consisting of tomato cultivars M82, Beijing 1, TS545, TS181, TS590.

55. The method of any one of claims 28-54, wherein the hybrid male plant is the wild tomato Solanum pimpinellifolium.

56. The method of any one of claims 28-55, wherein the male and / or female hybrid plants are grown in a phytotron based on LED lighting.

57. The method of claim 56, wherein the male and / or female hybrid plants are cultured under LED lighting and a day length of approximately 16 to 22 hours.

58. A method for automated pollination, comprising: detecting the presence of stigmas of crops to be pollinated; Acquire images of the crop stigma at two different positions, and obtain a three-dimensional location of the crop stigma based on matching feature points between the images; Based on the three-dimensional positioning of the crop stigma, a pollination device carrying pollen is driven to move in front of the crop stigma; as well as The pollination device is driven to move sequentially between a plurality of points in a set three-dimensional space including the three-dimensional positioning until the pollination device coats pollen onto the crop stigma.

59. The method of claim 58, comprising acquiring images of the crop stigma at the two different locations using movement of a camera.

60. The method according to claim 58, wherein when matching the feature points between the images, an image of a petal region including the crop stigma is cropped from the image, and the matching of the feature points is performed only based on the cropped image.

61. The method according to any one of claims 58 to 60, wherein the three-dimensional space is set as a stereoscopic space that covers a maximum positioning error of the three-dimensional positioning of the crop stigma in a three-dimensional coordinate system obtained by matching the feature points.

62. The method according to any one of claims 58 to 60, wherein the pollination device first moves to the three-dimensional location of the crop stigma obtained by matching the feature points, and gradually moves to the outer contour of the three-dimensional space through the multiple points from near to far.

63. The method according to claim 62, wherein the three-dimensional space is constructed as a cylinder, the center of mass of the cylinder is the three-dimensional position of the crop stigma obtained by matching the feature points, and the pollination device is driven to move sequentially along multiple points on the circular spiral line in at least two circular cross-sections of the cylinder until the pollination device coats pollen onto the crop stigma.

64. The method of claim 63, wherein the span between adjacent said annular spirals is determined by said maximum positioning error.

65. According to the method described in any one of claims 58 to 60, an image including the crop stigma and the pollination device is obtained each time the pollination device moves to one of the multiple points, and whether the crop stigma and the pollination device are in contact is determined based on their relative positions in the image, and when it is determined that the crop stigma and the pollination device are in contact twice in a row, it is determined that the pollination device has applied pollen to the crop stigma.

66. A system (100) for automated pollination, comprising: a movable base (110); a robotic arm assembly (120) mounted on the base (110); a pollination device (130), a camera (140), and a pollen box (150) detachably attached to the robot arm assembly (120); and A controller (160) configured to: controlling the camera (140) to acquire an image of a crop to detect the presence of a stigma of the crop to be pollinated; Controlling the camera (140) to acquire images of the crop stigma at two different positions, and acquiring a three-dimensional position of the crop stigma based on matching of feature points between the images; Based on the three-dimensional positioning of the crop stigma, the base (110) is driven to move, so that the pollination device (130) and the pollen box (150) carried by the mechanical arm assembly (120) are moved to the front of the crop stigma; as well as The pollination device (130) is driven to move sequentially between a plurality of points in a set three-dimensional space including the three-dimensional positioning until the pollination device (130) applies the pollen obtained from the pollen box (150) to the crop stigma.

67. A machine-readable storage medium storing executable instructions which, when executed by a processor module, implement the method according to any one of claims 58 to 65.

68. A method for processing an image containing flowers, comprising: A) Input image containing flowers; B) Send the input image to the backbone network for feature extraction to obtain a feature map; C) sending the obtained feature map to the detection branch to obtain prediction coefficients, and sending the obtained feature map to the segmentation branch to obtain a prototype mask, wherein the obtained prediction coefficients include coefficients representing the flower orientation; and D) The obtained prototype mask is combined with the prediction coefficients, and the output result is obtained together with the coefficient representing the flower orientation through the cropping and threshold modules.

69. The method according to claim 68, characterized in that In step A), the image is preprocessed, wherein the size of the input image is adjusted to W×H×3.

70. The method according to claim 68 or 69, characterized in that Step B) comprises: b1) performing feature extraction on the input image to obtain a feature backbone consisting of feature maps c1 to c5; and b2) Obtain a Feature Pyramid Network (FPN) consisting of feature maps p3 to p7 from the obtained feature maps c1 to c5.

71. The method according to claim 70, characterized in that In step b1), five feature maps c1 to c5 with decreasing sizes are generated through a fully convolutional network (FCN).

72. The method according to claim 70 or 71, characterized in that In step b2), the feature map c5 with the smallest size is passed through a convolution layer to obtain the feature map p5; the feature map p5 is amplified by bilinear interpolation once and added to the convolved feature map c4 to obtain the feature map p4; the feature map p4 is amplified by bilinear interpolation once and added to the convolved feature map c3 to obtain the feature map p3; the feature map p5 is convolved to obtain the feature map p6; the feature map p6 is convolved to obtain the feature map p7.

73. The method according to any one of claims 68 to 72, characterized in that Step C) comprises: c1) Send all feature maps p3 to p7 of the Feature Pyramid Network (FPN) to the detection branch to obtain the coefficient c+b+o for each anchor box, where the detection branch includes a prediction head and a non-maximum suppression module (NMS). The coefficient c represents the classification confidence, the coefficient b represents the bounding box regression, and the coefficient o represents the flower orientation. c2) At the same time as step c1), the maximum feature map p3 of the feature pyramid network (FPN) is sent to the segmentation branch parallel to the detection branch, where m prototype masks are obtained.

74. The method according to any one of claims 68 to 72, characterized in that The flower orientation refers to the direction of the flower stigma, which includes the following five categories: left, right, front, top and bottom.

75. The method according to any one of claims 68 to 74, characterized in that The loss function is expressed by the following formula (1): L = w cls ·L cls +w box ·L box +w mask ·L mask +w orient ·L orient (1) in, L cls Represents the classification confidence loss of whether the target category is a flower, and its weight coefficient is w cls , L box Represents the bounding box regression loss, whose weight coefficient is w box , L mask Represents the mask loss, and its weight coefficient is w mask , L orient Represents the flower orientation loss, and its weight coefficient is w orient .

76. The method according to claim 75, characterized in that Flower orientation loss L orient It is represented by the following formula (2): Among them, θ represents the direction angle of the flower stigma frame, for example, its range is set to 77. The method according to any one of claims 68 to 76, characterized in that The output result obtained in step D) is (x, y, w, h, c, m, o), where x, y, w, h represent the position, width, and height of the flower box, c represents the flower classification confidence, m represents the number of prototype masks, and o represents the flower's top, bottom, left, right, and front directions.