A blueberry picking and sorting robot autonomous navigation picking and sorting method and system

Abstract

The application relates to the field of agricultural robots and discloses a blueberry picking and sorting robot autonomous navigation picking and sorting method and system, wherein the plant distribution and spatial structure are acquired through an environment perception module, a density self-adaptive path and a resource pre-allocation strategy are generated by a global task planner, a picking and sorting module cooperates according to the strategy, a material buffer module dynamically adjusts the two end rates according to a load threshold, and a multi-mode feedback data driven double closed loop PID algorithm realizes millisecond level rate synchronization. The system integrates a laser radar, a stereo vision and a near infrared sensor, is equipped with a flexible clamping mechanical arm, a multi-stage screen sorting mechanism and a conical container buffer module, supports abnormal self-correction and multi-machine cooperative load redistribution. Through a dynamic coupling control architecture, the application improves the working efficiency, reduces the fruit damage rate and guarantees the continuous stable operation of the system for more than 8 hours.

Description

Technical Field

[0001] This invention belongs to the field of agricultural robots, specifically relating to an autonomous navigation method and system for blueberry picking and sorting robots. Background Technology

[0002] With the acceleration of automation and intelligentization in modern agriculture, fruit and vegetable harvesting and sorting operations are gradually shifting from manual labor to robotic systems. This is especially true in high-value-added berry sectors such as blueberries, where the requirements for harvesting efficiency, fruit damage prevention, and sorting accuracy are becoming increasingly stringent. Currently, most mainstream harvesting robots adopt a modular architecture, controlling the harvesting execution mechanism and sorting processing unit as independent subsystems, lacking real-time status perception and dynamic coordination mechanisms between the two.

[0003] The picking end performs batch picking based on preset paths and visual recognition results, while the sorting end grades the fruits on the conveyor belt at a fixed rate. The entire process relies on static parameter configuration and cannot be dynamically adjusted according to the fruit ripeness distribution, the workload of the robotic arm, or the congestion of the sorting channel. This results in frequent resource mismatches when the system is running under high load, such as picking too fast, causing sorting backlog or sorting idling while waiting for picking supply. This seriously restricts the overall throughput and the commercialization rate of the fruit.

[0004] Among them, blueberry picking and sorting operations are highly sensitive to fruit integrity and maturity grading. The core challenge lies in how to achieve seamless coordination among the three aspects of adaptive control of picking intensity, real-time determination of fruit maturity, and dynamic allocation of sorting paths in high-speed continuous operation.

[0005] While existing technologies have made partial breakthroughs in single aspects such as the maturity of visual recognition or the compliant control of robotic arms, they have not yet built a closed-loop feedback system covering the entire chain of picking perception, central decision-making, and sorting response. In particular, they lack an intelligent control mechanism that can adjust picking priority and execution parameters in reverse based on the real-time load status of the sorting unit. This results in insufficient robustness and efficiency bottlenecks in the system when facing fluctuations in the orchard environment, uneven fruit distribution, or differences in equipment performance.

[0006] In existing technologies, harvesting robotic arms generally employ open-loop force control strategies, making it difficult to dynamically adjust gripping force based on fruit texture, easily causing skin damage. While vision systems can identify ripeness, the data is not linked to the sorting unit's status, resulting in highly ripe fruits not being prioritized and remaining on the conveyor belt. Sorting mechanisms switch channels according to a fixed time sequence, unable to adaptively adjust channel speed based on upstream harvesting rhythm and fruit attribute distribution. These shortcomings are amplified dramatically in large-scale continuous operation scenarios, not only reducing effective output per unit time but also causing a decrease in marketable fruit rate due to fruit crushing, retention, or missorting, forming a key technological bottleneck restricting the automation upgrade of blueberry harvesting.

[0007] Therefore, there is an urgent need for an autonomous navigation picking and sorting method and system that can achieve deep collaboration among the three units of picking, sorting and central control, in order to break through the limitations of static control in the existing architecture and build a new paradigm of intelligent operation with dynamic balance and closed-loop optimization. Summary of the Invention

[0008] This invention provides an autonomous navigation method and system for blueberry picking and sorting robots. By constructing a dynamic collaborative control architecture with dual modules for picking and sorting, it achieves real-time closed-loop matching between picking and sorting operation rates, eliminating the problem of operational flow imbalance caused by independent module control. The system, with a global task planner at its core, integrates an environmental perception module, a picking execution module, a sorting execution module, a material buffer module, and a multimodal state feedback module. Through establishing a picking and sorting flow rate coupling model and a dynamic load balancing algorithm, it ensures that the output rate at the picking end and the processing rate at the sorting end remain synchronized in the temporal dimension, avoiding picking overload leading to sorting congestion or sorting idleness causing picking stagnation.

[0009] In a preferred embodiment of the present invention, the autonomous navigation picking and sorting method of the blueberry picking and sorting robot includes:

[0010] The environmental perception module acquires three-dimensional spatial structure information of the work area and blueberry plant distribution density information.

[0011] The global task planner generates a picking path sequence and a sorting resource pre-allocation strategy based on plant distribution density information.

[0012] The picking execution module performs fruit positioning, clamping and picking, and fruit transfer operations according to the picking path sequence, while recording the number of fruits picked per unit time.

[0013] The sorting execution module performs fruit size grading, maturity identification, and classification and packing operations according to the sorting resource pre-allocation strategy, while recording the number of fruits sorted per unit time.

[0014] An adjustable fruit storage area is established between the picking execution module and the sorting execution module through a material buffer module. The storage area has upper and lower thresholds. When the number of fruits in the storage area reaches the upper threshold, a deceleration command is sent to the picking execution module. When the number of fruits in the storage area is lower than the lower threshold, a deceleration command is sent to the sorting execution module.

[0015] The multimodal state feedback module collects data in real time, including the output rate of the picking execution module, the processing rate of the sorting execution module, the current load of the temporary storage area, the density of remaining fruits on the plant, and the remaining battery power of the robot. This data is then input into a dynamic load balancing algorithm to dynamically adjust the clamping frequency of the picking execution module and the conveyor belt speed of the sorting execution module, ensuring that the difference between the picking output rate and the sorting processing rate remains within a preset tolerance range.

[0016] In a preferred embodiment of the present invention, the environmental perception module includes a lidar sensor, a stereo vision camera, and a near-infrared spectral sensor. The lidar sensor is mounted on a rotating gimbal on top of the robot and is used to scan the work area to generate a point cloud map and identify the position of the plant trunk and the width of the aisles between rows. The stereo vision camera is mounted in front of the end effector of the robotic arm and is used to perform three-dimensional spatial positioning of individual blueberry plants before harvesting, and output the center coordinates of the fruit and the harvesting and holding angle. The near-infrared spectral sensor is mounted above the entrance of the sorting conveyor belt and is used to collect the surface reflectance spectrum of the fruit before it enters the sorting process to determine the sugar content and ripeness level of the fruit.

[0017] In a preferred embodiment of the present invention, the global task planner incorporates a plant density adaptive path generation algorithm. This algorithm divides the work area into high-density, medium-density, and low-density zones based on the number of plants per unit area in the LiDAR point cloud map. In the high-density zone, a serpentine reciprocating path is used, with the robotic arm's working radius covering two adjacent rows of plants. In the medium-density zone, a single-row straight path is used, with the robotic arm's working radius covering a single row of plants. In the low-density zone, a jump-point-to-point path is used, with the robotic arm only initiating the picking action when a cluster of fruits is detected. Simultaneously, sorting resources are pre-allocated according to the area density level. The sorting conveyor belt speed in the high-density zone is increased to 120% of the base speed, the medium-density zone maintains the base speed, and the low-density zone is reduced to 80% of the base speed.

[0018] In a preferred embodiment of the present invention, the harvesting execution module includes a six-degree-of-freedom robotic arm, a flexible gripping end effector, and a fruit transfer conduit. The base of the six-degree-of-freedom robotic arm is fixed to the robot chassis, and the flexible gripping end effector is installed at its end. The flexible gripping end effector consists of two symmetrically arranged silicone finger pads, with pressure sensors embedded in the finger pads. The gripping force is controlled within the range of 0.5N to 1.5N to avoid damage to the fruit skin. The fruit transfer conduit is a polytetrafluoroethylene pipe with a smooth inner wall and an inclination angle of 45°. It connects the gripping end to the inlet of the material buffer module, allowing the fruit to slide off without damage by gravity.

[0019] In a preferred embodiment of the present invention, the sorting execution module includes a vibrating feeder, a size grading screen group, a maturity recognition camera array, and a multi-channel classification guide plate. The vibrating feeder is located below the outlet of the material buffer module and generates directional vibration through an eccentric wheel drive, causing the fruit to spread evenly into the size grading screen group. The size grading screen group consists of three layers of parallel circular aperture screens: the first layer has an aperture of 12mm, the second layer has an aperture of 10mm, and the third layer has an aperture of 8mm, corresponding to large, medium, and small fruit grading, respectively. The maturity recognition camera array consists of three high-resolution color cameras that simultaneously capture the color distribution on the fruit surface from three perspectives: top, left, and right, and combine this with near-infrared spectral data to comprehensively determine the maturity level. The multi-channel classification guide plate is located at the end of the screen group and drives a pneumatic pusher based on the size and maturity determination results to guide the fruit into the corresponding grade collection box.

[0020] In a preferred embodiment of the present invention, the material buffer module includes a conical funnel container, a weighing sensor array, and a liquid level photoelectric detector. The upper opening diameter of the conical funnel container is 30cm, the lower outlet diameter is 8cm, and the internal volume is 5L. The weighing sensor array consists of four three-point weighing units, which are respectively installed at the four corners of the bottom of the funnel container to output the total weight of the container in real time. The liquid level photoelectric detector consists of three pairs of infrared transmitting and receiving tubes, which are equidistantly arranged along the height of the container and correspond to the lower threshold, the reference threshold, and the upper threshold, respectively. When the lower threshold tube is not blocked, the buffer area is determined to be in a low-load state; when the upper threshold tube is blocked, the buffer area is determined to be in a high-load state.

[0021] In a preferred embodiment of the present invention, the multimodal state feedback module includes a picking rate counter, a sorting rate counter, a plant remaining density estimator, and a power monitoring unit. The picking rate counter calculates the number of fruits picked per unit time using a robotic arm end-effector position sensor and a gripping action trigger signal. The sorting rate counter counts the number of fruits entering the sorting process per unit time using a photoelectric through-beam sensor at the inlet of the vibrating feeder. The plant remaining density estimator calculates the proportion of space occupied by the picked fruits based on the difference in point cloud volume of the same plant area before and after picking by a stereo vision camera, and inversely infers the density of the remaining fruits. The power monitoring unit calculates the remaining available power of the robot in real time using the current integration method. When the power is below 20%, a low-power mode is triggered, reducing the robotic arm's operating speed and the sensor sampling frequency.

[0022] In a preferred embodiment of the present invention, the dynamic load balancing algorithm adopts a dual-loop proportional-integral-derivative (PID) control structure. The outer loop controls the output rate of the picking execution module, and the inner loop controls the processing rate of the sorting execution module. The outer loop setpoint is the current processing rate of the sorting execution module, and the feedback value is the actual output rate of the picking execution module. After the error signal is tuned with a proportional coefficient of 0.8, an integral time constant of 5s, and a derivative time constant of 1s, the picking clamping frequency adjustment is output. The inner loop setpoint is the current output rate of the picking execution module, and the feedback value is the actual processing rate of the sorting execution module. After the error signal is tuned with a proportional coefficient of 0.6, an integral time constant of 3s, and a derivative time constant of 0.5s, the sorting conveyor speed adjustment is output. The two control loops share the load status of the material buffer module as a feedforward compensation signal. When the load is higher than the benchmark threshold, the outer loop proportional coefficient increases by 20%, and the inner loop proportional coefficient decreases by 15%, accelerating sorting and suppressing picking. When the load is lower than the benchmark threshold, the outer loop proportional coefficient decreases by 15%, and the inner loop proportional coefficient increases by 20%, accelerating picking and suppressing sorting.

[0023] In a preferred embodiment of the present invention, the blueberry picking and sorting robot autonomous navigation picking and sorting system includes a robot mobile chassis, a central control unit, a power management module, a wireless communication module, and a human-machine interface. The robot mobile chassis adopts a four-wheel independent drive structure, with each drive wheel equipped with a brushless DC motor and a planetary reducer, enabling in-situ turning and diagonal movement. The central control unit adopts a multi-core heterogeneous processor architecture, including a real-time control core for path planning, a graphics computing core for image processing, and a general-purpose computing core for algorithm execution. The power management module integrates a lithium battery pack and a bidirectional DC-DC converter, supporting dynamic adjustment of the power supply voltage of each module during operation. The wireless communication module supports dual-mode communication of 5G mobile communication technology and low-power Bluetooth, used to receive remote task commands and upload operation status data. The human-machine interface is a seven-inch resistive touch screen, displaying a map of the current operation area, a picking and sorting rate curve, buffer load status, and remaining battery percentage.

[0024] In a preferred embodiment of the present invention, the system performs a self-test process upon startup, sequentially detecting the integrity of the LiDAR point cloud generation, the zero-point position of the encoders at each joint of the robotic arm, the opening and closing stroke of the flexible gripper end effector, the amplitude stability of the vibrating feeder, the cleanliness of the size grading screen aperture, the white balance calibration status of the maturity recognition camera, the zero-point drift of the buffer weighing sensor, and the voltage sampling accuracy of the power monitoring unit. After passing the self-test, the system enters a standby state, waiting to receive instructions on the coordinates of the work area and the task priority. Upon receiving the instructions, the global task planner generates an initial path and resource allocation scheme, the environmental perception module begins continuous scanning, the picking and sorting execution modules enter a preheating standby mode, the material buffer module clears and resets the photodetector, the multimodal status feedback module starts data acquisition, and the dynamic load balancing algorithm initializes the control parameters.

[0025] In a preferred embodiment of the present invention, the system implements an anomaly handling mechanism during operation. When the picking execution module fails to grasp the fruit three times in a row, it is determined that the fruit positioning error exceeds the limit, the current plant operation is paused, and the three-dimensional reconstruction by the stereo vision camera is re-executed. When the sorting execution module determines that five fruits are immature in a row, it is determined that the maturity recognition model is drifting, the sorting process is paused, and the near-infrared spectral sensor is triggered to re-collect standard samples for online model correction. When the data of the buffer area weighing sensor and the photoelectric detector are inconsistent by more than 10%, it is determined that the fruit accumulation pattern is abnormal, the vibration motor is started to perform pulse disturbance on the buffer area to redistribute the fruits. When the remaining battery power of the robot is detected to be less than 10%, the current operation is forcibly terminated, the autonomous return program is started, and the robot returns to the charging station along the shortest path.

[0026] In a preferred embodiment of the present invention, the system supports a multi-robot collaborative operation mode. The central control unit receives the coordinates of the work area and the current load status broadcast by the neighboring robots through the wireless communication module. When it detects that the plant density in its own work area is lower than the threshold and the load in the buffer area of ​​the neighboring robot is higher than the upper limit, it actively sends a support request to the neighboring robot. After confirmation by the central control unit of the other party, it adjusts its own global task planner path and goes to the neighboring area to assist in harvesting. The harvested fruits are directly input into the buffer area of ​​the other party through a dedicated transfer conduit, realizing cross-robot load redistribution.

[0027] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0028] This invention completely solves the problem of work flow imbalance caused by independent module control in existing systems by constructing a dynamic collaborative control architecture for both picking and sorting modules.

[0029] The introduction of the material buffer module provides physical-level rate buffering capability. Combined with the dual closed-loop control structure of the dynamic load balancing algorithm, it realizes millisecond-level synchronous adjustment of the picking output rate and sorting processing rate, effectively avoiding sorting congestion and picking stagnation.

[0030] The global task planner adaptively adjusts the path and resource allocation based on the plant density, enabling the robot to maintain optimal operating efficiency in areas with different densities.

[0031] The real-time data provided by the multimodal state feedback module supports the accurate decision-making of the control algorithm, and the anomaly handling mechanism ensures the robustness of the system in complex orchard environments.

[0032] The multi-robot collaborative operation mode further improves the resource utilization and task flexibility of the overall operation system. Real-world test data shows that, in high-density blueberry orchard operations, compared to traditional independent control architectures, this invention improves overall picking and sorting efficiency, reduces fruit damage rate, and extends the system's continuous trouble-free operation time to over 8 hours. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall technical architecture of the blueberry picking and sorting robot autonomous navigation picking and sorting method and system proposed in this invention;

[0034] Figure 2 This is a schematic diagram of the core principle framework of the dual-module dynamic collaborative control architecture for picking and sorting in this invention;

[0035] Figure 3 This is a simulation diagram of the autonomous navigation path of the blueberry picking and sorting robot in this invention;

[0036] Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the picking execution module, the sorting execution module and the material buffer module in this invention; Detailed Implementation

[0037] Please refer to Figure 1-4This invention provides an autonomous navigation method and system for blueberry picking and sorting robots. Its core lies in constructing a dynamic collaborative control architecture with two modules: picking and sorting. Through the organic integration of a global task planner, an environmental perception module, a picking execution module, a sorting execution module, a material buffer module, and a multimodal status feedback module, real-time closed-loop matching of picking and sorting operation rates is achieved. This architecture completely eliminates the problem of operational flow imbalance caused by independent module control, ensuring that the output rate at the picking end and the processing rate at the sorting end remain synchronized in the temporal dimension, avoiding picking overload leading to sorting congestion or sorting idleness leading to picking stagnation. The entire system uses a central control unit as the computing hub, coordinating data acquisition, command issuance, and status feedback from various functional modules to ensure that the robot can stably, efficiently, and with low loss complete the entire process in complex orchard environments.

[0038] In this embodiment, the autonomous navigation picking and sorting method of the blueberry picking and sorting robot includes: acquiring three-dimensional spatial structure information and blueberry plant distribution density information of the work area through an environmental perception module; generating a picking path sequence and a sorting resource pre-allocation strategy based on the plant distribution density information through a global task planner; performing fruit positioning, clamping and picking, and fruit transfer operations according to the picking path sequence through a picking execution module, while recording the number of fruits picked per unit time; and performing fruit size grading, maturity identification, and sorting and boxing operations according to the sorting resource pre-allocation strategy through a sorting execution module, while recording the number of fruits sorted per unit time.

[0039] An adjustable-capacity fruit storage area is established between the picking and sorting execution modules via a material buffer module. This storage area has upper and lower thresholds. When the number of fruits in the storage area reaches the upper threshold, a deceleration command is sent to the picking execution module; when the number of fruits in the storage area falls below the lower threshold, a deceleration command is sent to the sorting execution module. A multimodal status feedback module collects real-time data on the output rate of the picking execution module, the processing rate of the sorting execution module, the current load of the storage area, the density of remaining fruits on the plant, and the remaining battery power of the robot. This data is then input into a dynamic load balancing algorithm to dynamically adjust the clamping frequency of the picking execution module and the conveyor belt speed of the sorting execution module, ensuring that the difference between the picking output rate and the sorting processing rate remains within a preset tolerance range.

[0040] The environmental perception module consists of a lidar sensor, a stereo vision camera, and a near-infrared spectral sensor. These three components complement each other through spatiotemporal registration and multi-source data fusion algorithms. The specific fusion process is as follows:

[0041] (I) Hardware Installation and Spatiotemporal Calibration

[0042] Spatial calibration: A global coordinate system is established with the center of the robot chassis as the origin. The installation positions of the LiDAR, stereo vision camera, and near-infrared spectral sensor are obtained through the calibration board to obtain their extrinsic parameters in the global coordinate system, namely the translation matrix T and the rotation matrix R, to ensure that the data collected by the three are mapped to the same spatial dimension.

[0043] Time synchronization: A hardware trigger signal is used to uniformly control the sampling time of the three sensors with a 100Hz synchronization pulse, and the time deviation is controlled within ±1ms to avoid data misalignment caused by sampling delay.

[0044] (II) Specific process of data fusion

[0045] Spatial Fusion of LiDAR and Stereo Vision

[0046] The lidar scans the work area 360° at an angular velocity of 10 revolutions per second, generating a 3D point cloud map containing the location of the plant trunks, the width of the aisles between rows, and the height of the ground undulations. The point cloud data is processed by voxel filtering: the grid size is 5cm×5cm×5cm and the plane is segmented using the RANSAC algorithm: distance threshold of 0.05m and 1000 iterations. The cylindrical model of the plant trunk is extracted, the radius fitting error is ≤0.02m, the distance between adjacent trunks is calculated, and a heat map of plant distribution density is generated.

[0047] A stereo vision camera with a binocular baseline distance of 60mm performs local 3D reconstruction of the target plant area marked by LiDAR before the picking action begins. The disparity map is calculated using the SGBM (semi-global block matching) algorithm. Combined with camera intrinsic parameters (focal length f=8mm, principal point coordinates (u0=640px, v0=480px)) and extrinsic parameters, the 3D coordinates (X, Y, Z) of the fruit center in the robotic arm's base coordinate system are output. The calculation formula is as follows:

[0048] ;

[0049] ;

[0050] ;

[0051] Where (u,v) are the pixel coordinates of the fruit.

[0052] d represents parallax, in pixels (px).

[0053] B represents the binocular baseline distance, which is 60mm.

[0054] f is the focal length, which is 8mm.

[0055] Meanwhile, the required rotation angle of the clamping end is calculated by the normal vector of the fruit surface, and the rotation angle θ around the Z-axis of the robotic arm is accurate to ±0.5°.

[0056] Fusion logic: LiDAR provides coarse positioning of the plant area, and stereo vision achieves precise positioning of the fruit within the area. When stereo vision detects that the fruit coordinates are outside the plant range marked by LiDAR, a recalibration process is triggered to ensure positioning accuracy.

[0057] Maturity fusion of near-infrared spectroscopy and visual features

[0058] A near-infrared spectral sensor with a light source wavelength range of 900nm to 1700nm and a sampling interval of 5ms is used to collect the surface reflectance spectrum curve of the fruit before it enters the sorting process. Three characteristic bands are selected: 1100nm (sugar content sensitive band), 1450nm (moisture sensitive band), and 1650nm (maturity sensitive band). The absorbance value is calculated as: A = log (1 / R), where R is the reflectance. The fruit sugar content value is output through a partial least squares regression model.

[0059] A maturity recognition camera array consists of three 5-megapixel cameras that capture RGB color features of the fruit surface. Three color feature parameters are extracted: H channel mean, S channel variance, and R / G ratio. A support vector machine classifier is used to preliminarily determine the maturity level: fully ripe, semi-ripe, or immature.

[0060] Fusion logic: A weighted voting mechanism is adopted, in which the near-infrared spectral sugar content value (weight 0.6) and the visual color feature (weight 0.4) jointly determine the final maturity level. When the two determination results are inconsistent, secondary sampling is triggered to re-acquire the spectrum and image to ensure that the maturity determination accuracy is ≥98%.

[0061] The global task planner incorporates a plant density-adaptive path generation algorithm. This algorithm first divides the work area into high-density, medium-density, and low-density zones based on the number of plants per unit area in the LiDAR point cloud map, with thresholds of more than 8 plants per square meter, 4 to 8 plants per square meter, and less than 4 plants per square meter, respectively. In the high-density zone, the planner generates a serpentine reciprocating path with a turning point spacing of 1.5 meters, ensuring the robotic arm's working radius covers two adjacent rows of plants to maximize the harvesting area covered in a single path. In the medium-density zone, the planner generates a single-row straight path extending along the centerline of the plant row, with the robotic arm's working radius covering a single row of plants to reduce unnecessary movement. In the low-density zone, the planner generates a jump-point-to-point path, only generating paths when the stereo vision camera detects a fruit cluster with a 3D volume greater than 50 m³. 3Only when the picking point coordinates are generated will the robotic arm initiate its gripping action upon reaching the picking point. Simultaneously, the planner pre-allocates sorting resources based on the area density level: in high-density areas, the sorting conveyor belt speed is increased to 120% of the baseline speed (0.3 m / s); in medium-density areas, the baseline speed is maintained; and in low-density areas, the speed is reduced to 80% of the baseline speed. This pre-allocation strategy ensures that sorting capacity and picking intensity are pre-matched spatially, reducing the burden of dynamic adjustment.

[0062] The harvesting execution module consists of a six-degree-of-freedom robotic arm, a flexible gripping end effector, and a fruit transfer conduit. The six-degree-of-freedom robotic arm base is fixed to the front of the robot chassis, and each joint uses harmonic reducer transmission, achieving a repeatability accuracy of ±0.1mm. The flexible gripping end effector consists of two symmetrically arranged silicone finger pads with a radius of curvature of 15mm, embedded with a thin-film pressure sensor at a sampling frequency of 1000Hz, providing real-time feedback of the gripping force. The gripping force control algorithm employs a proportional-integral closed-loop structure, with a set value of 1N and an allowable fluctuation range of 0.5N to 1.5N, ensuring no indentations or cracks on the fruit skin during gripping. The fruit transfer conduit is a smooth-walled polytetrafluoroethylene (PTFE) pipe with an inner diameter of 25mm and a fixed tilt angle of 45°. Its upper end connects to the outlet below the gripping end effector, and its lower end connects to the inlet of the material buffer module. The fruit slides down the conduit by gravity in less than 0.5s, without collision or tumbling, avoiding secondary damage. After each gripping action is completed, the robotic arm control program automatically raises the end effector to a safe height and then moves along a preset path to the next picking point, ensuring that there is no interference from branches and leaves during the movement.

[0063] The sorting execution module consists of a vibrating feeder, a size grading screen group, a maturity recognition camera array, and a multi-channel classification guide plate. The vibrating feeder, measuring 400mm × 300mm, is positioned directly below the material buffer module outlet. Driven by an eccentric wheel, it vibrates at a frequency of 20Hz with an amplitude of 3mm, ensuring the fruits are evenly spread and arranged in a single layer before entering the size grading screen group. The size grading screen group comprises three layers of parallel circular mesh screens made of stainless steel. The first layer has a 12mm aperture for separating large fruits; the second layer has a 10mm aperture for separating medium fruits; and the third layer has an 8mm aperture for separating small fruits. Fruits that fail to pass through any screen are considered oversized or impurities and are discharged from the final waste outlet. The maturity recognition camera array consists of three 5-megapixel color cameras, mounted above, to the left, and to the right of the screen group, with shooting angles of 90° to each other. Exposure is triggered synchronously to acquire the red, blue, and green color distribution on the fruit surface.

[0064] Combining sugar content values ​​from near-infrared spectroscopy sensors, a support vector machine classifier is used to determine maturity levels, categorizing them into fully ripe, semi-ripe, and immature. A multi-channel classification guide plate is located at the end of the sieve assembly and is driven by six pneumatic pushers, each corresponding to a collection box with a specific size and maturity combination. After the fruit passes through the sieve and its maturity is determined, the guide plate adjusts its angle within 0.1 seconds, guiding the fruit into the corresponding collection box, ensuring a classification accuracy of no less than 98%.

[0065] The material buffer module consists of a conical funnel container, a weighing sensor array, and a liquid level photoelectric detector. The conical funnel container has an upper opening diameter of 30cm, a lower outlet diameter of 8cm, a total height of 40cm, an internal volume of 5L, and is made of food-grade polypropylene. The weighing sensor array consists of four three-point weighing units, installed at the four corners of the bottom of the funnel container. Each weighing unit has a weighing range of 5kg and an accuracy of ±5g. The four signals are summed by a circuit to output the total weight of the container, with a sampling frequency of 10Hz. The liquid level photoelectric detector consists of three pairs of infrared transmitting and receiving tubes, equidistantly arranged along the height of the container at 10cm intervals, corresponding to the lower threshold, reference threshold, and upper threshold, respectively.

[0066] The lower threshold is located at 20% of the container height, the baseline threshold at 50%, and the upper threshold at 80%. When the lower threshold tube is not obstructed, the buffer area is determined to be in a low-load state, triggering a deceleration command for the sorting execution module; when the upper threshold tube is obstructed, the buffer area is determined to be in a high-load state, triggering a deceleration command for the picking execution module. Weighing data and photoelectric detection data are cross-verified; if the discrepancy between the two exceeds 10%, an exception handling mechanism is activated.

[0067] The multimodal state feedback module consists of a picking rate counter, a sorting rate counter, a plant remaining density estimator, and a power monitoring unit. The picking rate counter calculates the number of fruits picked per unit time by combining the position sensor at the end of the robotic arm with the gripping action trigger signal. Specifically, whenever the gripping end effector completes a closing action and detects a pressure value greater than 0.3N, the counter increments by 1, and a timestamp is recorded. The count value is calculated every 5 seconds, and the current picking rate is output as the number of fruits picked per second.

[0068] The sorting rate counter uses a photoelectric sensor at the inlet of the vibrating feeder to count the number of fruits entering the sorting process per unit time. The photoelectric sensor beam spacing is 25mm. Whenever the beam is blocked once and the duration is greater than 50ms, it is determined that a fruit has passed through, and the counter is incremented by 1. The count value is calculated every 5s, and the current sorting rate is output as the number of fruits sorted per second. The plant remaining density estimator calculates the proportion of space occupied by the harvested fruits based on the difference in point cloud volume of the same plant area before and after harvesting by a stereo vision camera. The specific process is as follows: before harvesting, a 3D scan of the target plant area is performed to calculate the total point cloud volume including the fruits;

[0069] After harvesting, scan the same area again and calculate the remaining point cloud volume; divide the difference between the two by the average volume of a single fruit, 10 cm³. 3 The number of harvested fruits is obtained; combined with the estimated total number of fruits on the original plant, the density of remaining fruits is deduced. The power monitoring unit calculates the robot's remaining available power in real time using the current integration method. The current sensor is connected in series to the positive output terminal of the lithium battery pack, with a sampling frequency of 100Hz. The cumulative discharge current is integrated over time, and combined with the battery's nominal capacity of 20A, the remaining power percentage is calculated. When the power is below 20%, a low-power mode is triggered, reducing the robotic arm's operating speed by 30% and the sensor sampling frequency by 50%.

[0070] The dynamic load balancing algorithm employs a dual-loop proportional-integral-derivative (PID) control structure. The outer loop controls the output rate of the picking execution module, while the inner loop controls the processing rate of the sorting execution module. Details are as follows:

[0071] (a) Error signal calculation method

[0072] Outer loop error, picking rate error E1: The outer loop setting value is the current processing rate V of the sorting execution module. s The feedback value is the actual output rate V of the picking execution module. p The formula for calculating the error signal is:

[0073] E;

[0074] Where k is the sampling period (k=1,2,..., sampling period T=0.1s), and an error integral term ∫E1(t)dt is introduced, with an integration time range of 0-5s and a differential term ΔE1 / Δt, with a differential time step of 0.05s, to avoid static error and rate abrupt changes.

[0075] Inner loop error, sorting rate error E2: The inner loop setting value is the current output rate V of the picking execution module. p The feedback value is the actual processing rate V of the sorting execution module. s The formula for calculating the error signal is:

[0076] ;

[0077] Similarly, by introducing integral and differential terms, we can ensure that the sorting rate closely follows the picking rate.

[0078] (ii) Control signal conversion process

[0079] Picking and clamping frequency adjustment conversion

[0080] The outer loop PID output is the picking and holding frequency adjustment Δf, calculated using the following formula:

[0081] ;

[0082] Among them, K p1 =0.8, K i1 =0.16, K i1 =K p1 / T i T i =5s、K d1 =0.08, K d1 =K p1 ×T d T d =0.1s.

[0083] The clamping frequency reference value is f0=1Hz. After adjustment, the actual clamping frequency is f(k)=f0+Δf(k), and the adjustment range is -50% to +50%.

[0084] Control signal conversion: f(k) is converted into a joint control signal for the robotic arm. The six-degree-of-freedom robotic arm is driven by a pulse width modulation (PWM) signal. The PWM frequency is proportional to the gripping frequency f(k), with a proportionality coefficient of 1000, that is, 1Hz corresponds to a 1000Hz PWM signal. At the same time, combined with the pressure feedback (0.5N to 1.5N) at the flexible gripping end, the PWM duty cycle (50% to 80%) is dynamically adjusted to avoid fruit damage.

[0085] Sorting conveyor belt speed adjustment conversion

[0086] The inner loop PID output is the speed adjustment amount Δv of the sorting conveyor belt, calculated using the following formula:

[0087] ;

[0088] Among them, K p2 =0.6, K i2 =0.2, K i2 =K p2 / T i T i=3s、K d2 =0.03, K d2 =K p2 ×T d T d =0.05s.

[0089] The reference speed of the conveyor belt is v0 = 0.3 m / s. The actual speed after adjustment is v(k) = v0 + Δv(k), and the adjustment range is -40% to +40%.

[0090] Control signal conversion: v(k) is converted into a conveyor belt motor control signal, and the corresponding speed command is output through the DC motor driver. The transmission ratio of the motor speed to the conveyor belt speed v(k) is 100. At the same time, the motor speed is fed back in real time through the photoelectric encoder to form a speed closed loop and ensure the speed control accuracy is ±0.01m / s.

[0091] The inner loop setpoint is the current output rate of the picking execution module, and the feedback value is the actual processing rate of the sorting execution module. The error signal, after being tuned with a proportional coefficient of 0.6, an integral time constant of 3s, and a derivative time constant of 0.5s, outputs the adjustment amount for the sorting conveyor belt speed. The base value for the conveyor belt speed is 0.3m / s, and the adjustment range is -41% to +40%, meaning the conveyor belt speed can be dynamically adjusted between 0.18m / s and 0.42m / s. The two control loops share the load status of the material buffer module as a feedforward compensation signal.

[0092] When the load exceeds the baseline threshold, the outer loop proportional coefficient increases by 20% to 0.96, and the inner loop proportional coefficient decreases by 15% to 0.51, accelerating sorting and suppressing picking. When the load falls below the baseline threshold, the outer loop proportional coefficient decreases by 15% to 0.68, and the inner loop proportional coefficient increases by 20% to 0.72, accelerating picking and suppressing sorting. This feedforward compensation mechanism improves system response speed by 40% and reduces overshoot by 60%.

[0093] The blueberry picking and sorting robot's autonomous navigation system includes a robot chassis, a central control unit, a power management module, a wireless communication module, and a human-machine interface. The robot chassis employs a four-wheel independent drive structure, with each drive wheel equipped with a 500W brushless DC motor and a planetary reducer with a reduction ratio of 12:1. The maximum travel speed is 1.5m / s, the minimum turning radius is 0, and it supports turning in place and diagonal movement in any direction.

[0094] The central control unit adopts a multi-core heterogeneous processor architecture, including a real-time control core with an operating frequency of 800MHz for executing path planning and motion control algorithms; a graphics computing core with an operating frequency of 1.2GHz for processing stereo vision and maturity recognition image data; and a general-purpose computing core with an operating frequency of 1.5GHz for executing dynamic load balancing algorithms and multimodal data fusion. The three cores exchange data via shared memory and message queues to ensure real-time synchronization of control commands and status feedback. The power management module integrates a 48V, 20Ah lithium battery pack and a bidirectional DC-DC converter, supporting stable voltages of 24V, 12V, and 5V for the robotic arm, sensors, and computing unit respectively, and dynamically adjusting the output current according to the load status, with a maximum output power of 1000W.

[0095] The wireless communication module supports dual-mode communication using 5G and Bluetooth Low Energy. 5G is used to receive remote task commands and upload work status data, with a communication range of up to 1km. Bluetooth Low Energy is used to broadcast load status to nearby robots, with a communication range of 50m. The human-machine interface is a 7-inch resistive touchscreen with a resolution of 800×80 pixels. The displayed content includes a map of the current work area, a picking and sorting rate curve, a bar chart of the buffer area load status, and a remaining battery percentage. It supports manual pause, resumption, or emergency stop of the operation by the operator.

[0096] The system performs a self-test process upon startup. The self-test process sequentially checks the integrity of the LiDAR point cloud generation: sending a 360° scan command, receiving point cloud data packets, and verifying that the number of points is greater than 100,000; checking the zero-point position of the encoders at each joint of the robotic arm: driving each joint to rotate to the mechanical zero point, reading the encoder values, and verifying that they are within the preset error range; and checking the opening and closing stroke of the flexible gripping end effector: driving the finger pad to complete three full-open and full-close cycles, verifying that the opening and closing angle is 90°, and that the zero-point drift of the pressure sensor is less than 0.05N.

[0097] To test the stability of the vibrating feeder amplitude: start the vibrating motor for 10 seconds, use an accelerometer to measure the amplitude of the feeder, and verify whether it is stable within 3mm ± 0.2mm; To test the cleanliness of the size grading screen aperture: pass standard steel balls with diameters of 12mm, 10mm, and 8mm through each layer of the screen in sequence, and verify whether the pass rate is 100%; To test the white balance calibration status of the maturity recognition camera: photograph a standard white calibration plate, calculate the ratio of the red, green, and blue channel averages, and verify whether it is between 0.95 and 1.05; To test the zero-point drift of the buffer area weighing sensor: empty the container, read the four weighing signals, and verify whether the sum is less than 20g; To test the voltage sampling accuracy of the power monitoring unit: input 48V with a standard voltage source, read the sampled value, and verify whether the error is less than 0.05%.

[0098] After the self-test passes, the system enters standby mode, and the touchscreen displays the word "Standby," awaiting instructions on the work area coordinates and task priority. Upon receiving the instructions, the global task planner generates an initial path and resource allocation scheme, and the environment awareness module begins continuous scanning, as detailed below:

[0099] (I) Key Steps and Parameters of Point Cloud Processing Algorithm

[0100] Voxel filtering: The raw point cloud acquired by the lidar is subjected to voxel filtering at a point cloud density of 100 points / cm³. The grid size is set to 5cm×5cm×5cm. By downsampling, key features of the point cloud are preserved, while redundant data is removed. The number of points in the cloud is reduced by 60% to 70%, which improves the efficiency of subsequent processing.

[0101] Plane segmentation, namely the RANSAC algorithm: used to separate ground and plant point clouds. The distance threshold is set to 0.05m, which is the maximum distance from a point to a plane. The number of iterations is 1000, the confidence level is 99.9%, and the equation of the ground plane is fitted to ax+by+cz+d=0. Point clouds that are more than 0.1m away from the ground plane are identified as plant point clouds, with a segmentation accuracy of ≥95%.

[0102] Clustering analysis: Euclidean clustering algorithm is used for plant point clouds. The clustering distance threshold is set to 0.3m, which is the maximum distance between adjacent points. The minimum number of cluster points is 500 to avoid noise interference. The maximum number of cluster points is 10,000, which is the upper limit of the number of individual plant point clouds. This achieves the segmentation of individual plants and provides a basis for subsequent density calculation.

[0103] (II) Fruit Localization Visual Algorithm

[0104] The stereo vision camera uses the SGBM algorithm to achieve 3D localization of the fruit. The specific steps and formulas are as follows:

[0105] Image preprocessing: The color images acquired by the left and right cameras, with a resolution of 1280×960px, are converted to grayscale, Gaussian filtered with a kernel size of 5×5 and a standard deviation of 1.2, and edge enhancement is performed to highlight the outline of the fruit.

[0106] Disparity calculation: Set SGBM algorithm parameters: minimum disparity 0, maximum disparity 64, block matching window size 11×11, P1 = 8×3×(window size) 2 =8×3×121=2904 (adjacent pixel parallax smoothing penalty), P2=32×3×(window size) 2 =32×3×121=11616 (disparity smoothing penalty for non-adjacent pixels), calculate the disparity map of the fruit region. .

[0107] 3D coordinate transformation: Combining camera intrinsic parameters, focal length f=8mm, principal point coordinates (u0=640px, v0=480px), and extrinsic parameters, the 3D coordinates (X,Y,Z) of the fruit center are calculated using the following formula:

[0108] ;

[0109] ;

[0110] ;

[0111] Let B be the center pixel coordinates of the fruit, and B be the binocular baseline distance (60mm). For parallax, coordinate calculation accuracy is ±0.005m.

[0112] (III) Details of the Maturity Classifier

[0113] The maturity classifier uses a near-infrared spectroscopy + RGB color feature fusion model, the details of which are as follows:

[0114] Spectral characteristics: Three key bands were selected:

[0115] 1100nm: Sugar content sensitive band, absorbance value is positively correlated with sugar content, correlation coefficient 0.85;

[0116] 1450nm: Moisture-sensitive band. Immature fruits have high water content, and their absorbance values ​​are significantly higher than those of mature fruits.

[0117] 1650nm: A wavelength sensitive to maturity. Mature fruits have a high anthocyanin content and lower absorbance than unripe fruits.

[0118] A sugar content prediction model was established using the PLSR model, with a cross-validation error ≤ 0.3 Brix.

[0119] Color features: Extracting HSV color space features from the RGB image of the fruit:

[0120] H-channel mean: H value range of 20° to 40° for mature fruit (red / purple), and H value range of 60° to 80° for immature fruit (green);

[0121] S channel variance S var Mature fruits have more uniform color saturation. var ≤0.05, S of immature fruit var ≥0.1;

[0122] R / G ratio: The R channel value of mature fruit is higher than that of the G channel, R / G ≥ 1.2, and the R / G ratio of immature fruit is ≤ 0.8.

[0123] Classifier training: An SVM classifier was used, with spectral features and color features as input vectors, 6 dimensions, and the kernel function set to RBF (radial basis function), penalty coefficient C=10, gamma=0.1. It was trained with 1000 labeled samples, 300 fully mature, 400 semi-mature, and 300 immature samples, achieving a classification accuracy of ≥98%.

[0124] The system implements an anomaly handling mechanism during operation. If the picking module fails to grasp the fruit three times consecutively, it is determined that the fruit positioning error exceeds the limit. The grasping failure criteria are: after the grasping action is completed, the pressure sensor value is less than 0.3N, and the stereo vision camera does not detect an update in the fruit's 3D coordinates. At this time, the system pauses the current plant operation, re-executes the 3D reconstruction with the stereo vision camera, adjusts the camera exposure parameters and focus distance, and recalculates the fruit coordinates. If the reconstruction still fails, the plant is marked as a difficult-to-harvest area, skipped, and its coordinates are recorded for subsequent manual processing. If the sorting module determines five consecutive fruits to be immature, it is determined that the maturity recognition model has drifted.

[0125] At this point, the system pauses the sorting process and triggers the near-infrared spectral sensor to re-collect standard samples. The standard samples are pre-calibrated fully ripe blueberries; their reflectance spectra are collected, compared with the built-in standard curve, the offset is calculated, the decision boundary parameters of the support vector machine classifier are updated, and sorting resumes after online model calibration. When a discrepancy of more than 10% is detected between the data from the buffer area weighing sensor and the photodetector, it is determined to be an abnormal fruit stacking pattern.

[0126] When the robot's remaining battery level is detected to be below 10%, the current operation is forcibly terminated, and the autonomous return-to-home procedure is initiated. The return-to-home procedure calls the LiDAR point cloud map and uses the A* algorithm to calculate the shortest path back to the charging station. The path avoids all plants and obstacles. The mobile chassis travels along the path at a speed of 0.8 m / s. Upon arrival at the charging station, it automatically docks with the charging interface and enters the charging state.

[0127] The system supports multi-robot collaborative operation. The central control unit broadcasts its own work area coordinates and current buffer load status every 10 seconds via a wireless communication module. When a robot detects that the plant density in its work area is less than 4 plants per square meter and the buffer load of a neighboring robot is higher than 80%, it proactively sends a support request data packet to the neighboring robot, including its own location, available capacity, and estimated arrival time.

[0128] Upon receiving a request, the central control unit of the neighboring robot assesses its own task urgency and path conflict risk. If the assessment passes, it replies with a confirmation command and adjusts its global task planner path, reserving an interface at the specified coordinates. This supports the robot adjusting its path to the nearby area; specific supplementary details are as follows:

[0129] (I) Mathematical formula for calculating plant density

[0130] The plant density ρ is calculated using the results of lidar point cloud clustering, as shown in the following formula:

[0131] ;

[0132] in:

[0133] N represents the number of plant clusters within the work area;

[0134] The average footprint of a single plant is calculated using the minimum bounding rectangle of the clustered point cloud. =Length × Width, where length and width are the two side lengths of the rectangle;

[0135] The area of ​​the task region is calculated using the coordinates of the region specified by the global task planner. If the region is rectangular, =Length × Width; for irregular areas, calculate using the polygon area formula.

[0136] Example: If the work area is 10m × 10m ( =100m 2 Clustering yielded N=50 plants, with an average area occupied by each plant. =0.8m 2 Therefore, ρ = 50 × 0.8 / 100 = 0.4 plants / m² 2 It was determined to be a low-density area.

[0137] (II) Path Planning Algorithm

[0138] The specific logic of the global task planner is as follows:

[0139] Raster map construction: The work area is divided into 0.5m×0.5m grids. Based on the LiDAR point cloud segmentation results, plant grids (grids with plants, impassable), passage grids (grids without plants, passable), and obstacle grids (grids such as stones, irrigation equipment, etc., impassable) are marked. The grid status update frequency is 1Hz.

[0140] Turning point generation

[0141] High-density areas ρ>0.8 plants / m² 2The serpentine reciprocating path is adopted, and the spacing between turning points is equal to the working radius of the robotic arm × 2. The working radius of the robotic arm is 1.5m, so the spacing between turning points is 3m. The coordinates (x, y) of the turning points satisfy: y = k × 3m (k = 0, 1, 2, ...). x moves back and forth between the left and right boundaries of the working area to ensure that the robotic arm covers two adjacent rows of plants.

[0142] Medium-density zone: 0.4 ≤ ρ ≤ 0.8 plants / m² 2 A single straight path is adopted, with turning points arranged along the center line of the plant row. The spacing between turning points is 5m, and the coordinates (x,y) satisfy: x = x coordinate of the center line of the row, y = k × 5m (k = 0, 1, 2, ...).

[0143] Low-density areas: ρ < 0.4 plants / m² 2 The system employs a jump-point-to-point path, with the turning point being the center coordinate of a single plant. The path only connects plants with fruit clusters, i.e., plants with ≥5 fruits, to avoid unnecessary movement.

[0144] Obstacle avoidance logic

[0145] Local obstacle avoidance: When the robot's mobile chassis detects an obstacle within 0.5m in front, the dynamic window method (DWA) is triggered. The velocity window is set to [0, 0.8m / s] and the angular velocity window is set to [-π / 4, π / 4] rad / s. The path cost within each window is evaluated, which is the minimum distance to the obstacle and the angular deviation to the target point. The movement command with the lowest cost is selected, and the obstacle avoidance response time is ≤0.1s.

[0146] Global replanning: If the obstacle persists for more than 5 seconds or the obstacle avoidance path deviates from the initial path by more than 1 meter, the global task planner will re-call the algorithm to update the path and turning points, ensuring that the robot avoids the obstacle while maintaining operational efficiency.

[0147] After the support robot arrives at the nearby area, the fruits picked by its harvesting module are directly transferred to the other robot's buffer area via a dedicated transfer conduit—a retractable flexible tube, 1m to 3m in length, 25mm inner diameter, and made of food-grade silicone. During the transfer, a weighing sensor records the number of fruits. During the support process, both robots share harvesting rate, sorting rate, and buffer load data via 5G communication, and uniformly adjust the gripping frequency and conveyor belt speed to achieve cross-robot load redistribution. After the collaborative operation is completed, both robots record the number of fruits, N. support With support time T support Used for subsequent task settlement, calculating workload and efficiency statistics based on the number of supported fruits, support efficiency = N support / T support .

[0148] This invention, through the organic integration of the aforementioned methods and systems, significantly improves the overall efficiency of picking and sorting, reduces fruit damage rate, and extends the system's continuous trouble-free operation time to over 8 hours compared to traditional independent control architectures in high-density blueberry orchard operations. This improvement stems from the dual safeguards of physical buffering and algorithmic adjustment: the material buffer module absorbs instantaneous rate fluctuations, and the dynamic load balancing algorithm achieves long-term rate matching; the global task planner's spatial adaptive strategy maximizes regional operational efficiency; multimodal state feedback provides precise control basis; the anomaly handling mechanism ensures system robustness; and the multi-robot collaborative mode improves overall resource utilization. All modules utilize industrial-grade components and a real-time operating system, ensuring stable operation in outdoor high-temperature, high-humidity, and dusty environments, meeting the stringent requirements of commercial agricultural production.

[0149] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0150] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for autonomous navigation and sorting of blueberries using a blueberry picking and sorting robot, characterized in that, include: The environmental perception module acquires three-dimensional spatial structure information of the work area and blueberry plant distribution density information. The global task planner generates a picking path sequence and a sorting resource pre-allocation strategy based on plant distribution density information, including: Based on the number of plants per unit area in the lidar point cloud map, the work area is divided into high-density area, medium-density area and low-density area. In high-density areas, a serpentine reciprocating path is used, and the robotic arm's operating radius covers two adjacent rows of plants. In medium-density areas, a single-row straight path is used, and the robotic arm's operating radius covers a single row of plants. In low-density areas, a jump-point-to-point path is used, and the robotic arm only starts the picking action when it detects a cluster of fruits. Sorting resources are pre-allocated according to the regional density level. The sorting conveyor belt speed in high-density areas is increased to 120% of the base speed, the medium-density areas maintain the base speed, and the low-density areas are reduced to 80% of the base speed. The picking execution module performs fruit positioning, clamping and picking, and fruit transfer operations according to the picking path sequence, while recording the number of fruits picked per unit time. The sorting execution module performs fruit size grading, maturity identification, and classification and packing operations according to the sorting resource pre-allocation strategy, while recording the number of fruits sorted per unit time. An adjustable fruit storage area is established between the picking execution module and the sorting execution module through a material buffer module. The storage area has upper and lower thresholds. When the number of fruits in the storage area reaches the upper threshold, a deceleration command is sent to the picking execution module. When the number of fruits in the storage area is lower than the lower threshold, a deceleration command is sent to the sorting execution module. The multimodal state feedback module collects data in real time, including the output rate of the picking execution module, the processing rate of the sorting execution module, the current load of the temporary storage area, the density of remaining fruits on the plant, and the remaining battery power of the robot. The output rate of the picking execution module, the processing rate of the sorting execution module, the current load of the temporary storage area, the density of remaining fruits on the plant, and the remaining battery power of the robot are input into a dynamic load balancing algorithm to dynamically adjust the clamping frequency of the picking execution module and the conveyor belt running speed of the sorting execution module, so that the difference between the picking output rate and the sorting processing rate is always maintained within a preset tolerance range, including: A dual-closed-loop proportional-integral-derivative control structure is adopted, with the outer loop controlling the output rate of the picking execution module and the inner loop controlling the processing rate of the sorting execution module. The outer ring setting value is the current processing rate of the sorting execution module, the feedback value is the actual output rate of the picking execution module, and the error signal is adjusted by a proportional coefficient of 0.8, an integral time constant of 5s, and a derivative time constant of 1s to output the clamping frequency adjustment amount. The inner loop setting value is the current output rate of the picking execution module, the feedback value is the actual processing rate of the sorting execution module, and the error signal is adjusted by a proportional coefficient of 0.6, an integral time constant of 3s, and a derivative time constant of 0.5s to output the conveyor belt speed adjustment amount. The load status of the shared material buffer module is used as a feedforward compensation signal. When the load is higher than the benchmark threshold, the outer loop proportional coefficient increases by 20% and the inner loop proportional coefficient decreases by 15%; when the load is lower than the benchmark threshold, the outer loop proportional coefficient decreases by 15% and the inner loop proportional coefficient increases by 20%.

2. The autonomous navigation picking and sorting method for blueberries using a picking and sorting robot according to claim 1, characterized in that, The environmental perception module acquires three-dimensional spatial structure information of the work area and blueberry plant distribution density information, including: A point cloud map is generated by scanning the work area using a lidar sensor to identify the location of the plant's main stem and the width of the aisles between rows. A stereo vision camera is used to locate a single blueberry plant in three-dimensional space and output the coordinates of the fruit center and the angle for picking and holding. The sugar content and ripeness grade of the fruit are determined by collecting the surface reflectance spectrum of the fruit using a near-infrared spectroscopy sensor.

3. The autonomous navigation picking and sorting method for blueberries using a picking and sorting robot according to claim 1, characterized in that, The harvesting execution module performs fruit positioning, clamping and harvesting, and fruit transfer operations according to the harvesting path sequence, including: A six-degree-of-freedom robotic arm drives a flexible gripper end effector to move to the center coordinate position of the fruit; Control the flexible gripping end effector to grip the fruit with a gripping force of 0.5N to 1.5N; The fruit is transferred through a fruit transfer conduit and falls by gravity to the inlet of the material buffer module.

4. The autonomous navigation picking and sorting method for blueberry picking and sorting robots according to claim 1, characterized in that, The sorting execution module performs fruit size grading, maturity identification, and sorting and packing operations based on the pre-allocation strategy of sorting resources, including: The fruit is evenly spread into the size grading screen group by a vibrating feeder. Large fruits (over 12mm), medium fruits (10mm to 12mm), and small fruits (8mm to 10mm) are screened separately using a three-layer circular aperture sieve. The maturity level of fruit is determined by combining a maturity recognition camera array with near-infrared spectral data. The fruit is guided into the corresponding grade collection box by a pneumatic push rod driven by a multi-channel classification guide plate.

5. The autonomous navigation picking and sorting method for blueberry picking and sorting robots according to claim 1, characterized in that, An adjustable-capacity temporary fruit storage area is established between the picking and sorting execution modules through a material buffer module, including: The harvesting module receives the fruit through a conical funnel container; The total weight of the container is output in real time via a weighing sensor array; Three pairs of infrared transmitter and receiver tubes are arranged at equal intervals along the height of the container using a liquid level photoelectric detector, corresponding to the lower threshold, the reference threshold, and the upper threshold, respectively. When the lower limit threshold is not blocked, the buffer is determined to be in a low-load state; When the upper limit threshold is blocked, the cache is determined to be in a high-load state.

6. The autonomous navigation picking and sorting method for blueberry picking and sorting robots according to claim 1, characterized in that, The multimodal state feedback module collects real-time data on the output rate of the picking execution module, the processing rate of the sorting execution module, the current load of the temporary storage area, the density of remaining fruits on the plants, and the remaining battery power of the robot, including: The number of fruits picked per unit time is calculated by combining the position sensor at the end of the robotic arm with the clamping action trigger signal. The number of fruits entering the sorting process per unit time is counted by photoelectric photoelectric sensors at the inlet of the vibrating feeder. The density of remaining fruit is inferred by using the difference in point cloud volume of the same plant area before and after harvesting with a stereo vision camera. The remaining available power of the robot is calculated in real time using the current integration method.

7. A blueberry picking and sorting robot autonomous navigation picking and sorting system, characterized in that, The method for implementing the autonomous navigation picking and sorting robot of blueberry picking and sorting as described in any one of claims 1 to 6 includes: The environmental perception module is used to acquire three-dimensional spatial structure information of the work area and blueberry plant distribution density information. A global task planner is used to generate picking path sequences and sorting resource pre-allocation strategies based on plant distribution density information. The harvesting execution module is used to perform fruit positioning, clamping and harvesting, and fruit transfer operations according to the harvesting path sequence, while recording the number of fruits harvested per unit time. The sorting execution module is used to perform fruit size grading, maturity identification and sorting and packing operations according to the sorting resource pre-allocation strategy, while recording the number of fruits sorted per unit time. The material buffer module is used to establish a fruit temporary storage area with adjustable capacity between the picking execution module and the sorting execution module. The temporary storage area has upper and lower thresholds. When the number of fruits in the temporary storage area reaches the upper threshold, a deceleration command is sent to the picking execution module. When the number of fruits in the temporary storage area is lower than the lower threshold, a deceleration command is sent to the sorting execution module. The multimodal status feedback module is used to collect data in real time, including the output rate of the picking execution module, the processing rate of the sorting execution module, the current load of the temporary storage area, the density of remaining fruits on the plant, and the remaining battery power of the robot. The dynamic load balancing algorithm module is used to receive the output rate of the picking execution module, the processing rate of the sorting execution module, the current load of the temporary storage area, the density of remaining fruits on the plant, and the remaining power of the robot, and dynamically adjust the clamping frequency of the picking execution module and the conveyor belt running speed of the sorting execution module so that the difference between the picking output rate and the sorting processing rate is always maintained within a preset tolerance range.

8. The autonomous navigation picking and sorting robot system for blueberries according to claim 7, characterized in that, The environmental perception module is used for: A point cloud map is generated by scanning the work area using a lidar sensor to identify the location of the plant's main stem and the width of the aisles between rows. A stereo vision camera is used to locate a single blueberry plant in three-dimensional space and output the coordinates of the fruit center and the angle for picking and holding. The sugar content and ripeness grade of the fruit are determined by collecting the surface reflectance spectrum of the fruit using a near-infrared spectroscopy sensor.

Images