Sorting machine grabbing control method based on artificial intelligence

Abstract

The invention relates to the technical field of intelligent sorting and robot grabbing control, and discloses a sorting machine grabbing control method based on artificial intelligence. According to the method, timestamps are written into camera frames, conveyor belt encoder data and tail end vacuum pressure and flow data under a unified space reference and a unified time reference and are aligned; a target instance set is obtained based on instance segmentation, the first model outputs a single-piece separability probability for gating, and when gating is not passed, a disturbance instruction is output and returned for collection; candidate points are generated for the target passing the gating, and geometric feasibility and time sequence accessibility judgment is executed; inputting the feasible candidates into a second model to output a capture success probability, and calculating score gating to select execution candidates; and a sealing index is calculated in the vacuum establishment window, vacuum setting and stripping parameters are adjusted in a self-adaptive mode, retry rollback and putting are completed in combination with criterion branches, and a structured result packet is output.

Description

Technical Field

[0001] This invention relates to the field of intelligent sorting and robot gripping control technology, specifically to a sorting machine gripping control method based on artificial intelligence. Background Technology

[0002] Automated sorting production lines are widely used in e-commerce warehousing, express delivery, and food and daily chemical warehouse sorting. A typical system usually consists of a conveyor belt feeding mechanism, a vision acquisition and recognition device, an industrial robot and end effector, a sorting machine controller, and a cycle control unit. In existing technologies, a common approach is to use industrial cameras or color and depth combined imaging to acquire image information of targets on the conveyor belt. Based on target detection or instance segmentation, the target contour or pose information is obtained, further generating gripping points or gripping postures, and driving the robot to perform gripping and delivery actions. To achieve consistency between visual coordinates and robot coordinates, it is usually necessary to perform camera intrinsic parameter calibration, hand-eye calibration, or establish a spatial mapping relationship between the camera coordinate system, robot coordinate system, and conveyor belt coordinate system using fixed markers. Simultaneously, the image and conveyor belt encoder signal are synchronized in time through the controller clock or acquisition board to meet the positioning and gripping requirements of the moving target.

[0003] However, in sorting scenarios involving flexible or deformable materials such as clothing, film packaging bags, and breathable mesh bags, targets on the conveyor belt are prone to folding, overlapping, entanglement, and partial obstruction. The graspable area and individual peelability of the target change rapidly with the contact relationship. In existing technologies, the generation of grasp points often relies primarily on single-frame geometric information or local surface features, which is insufficient to reliably reflect the "separability of individual components" of the target under the current contact state. This can easily lead to situations such as multiple components being picked up in a single adsorption, edge entanglement after adsorption causing placement deviations, or components slipping and falling off after grasping. Furthermore, the surface wrinkles, textures, and permeability differences of flexible materials can cause instability in the sealing state of vacuum adsorption. Existing systems often use "vacuum pressure reaching a set threshold" as the basis for grasp confirmation. Typically, only a single pressure threshold is used to determine whether adsorption is successful, making it difficult to distinguish between short-term achievement and sustained stable sealing. It is also difficult to use the pressure and flow rate changes during vacuum establishment to identify leakage, air leakage, or partial non-adhesion, which can easily lead to misjudgments and unnecessary retries.

[0004] On the other hand, sorting machines typically have fixed cycle times and loading window constraints. The time it takes for the target to reach the sorting port as it moves along the conveyor belt and the time required for the robot to arrive and complete the gripping process both vary with speed and position. In existing technologies, cycle time window constraints are often handled through external scheduling rules or simple time margin judgments, without forming a unified mechanism that incorporates "window reachability" into the candidate gripping feasibility domain determination. This leads to phenomena such as window misalignment, abandoned gripping, frequent speed reduction, or temporary changes to the gripping path when the window is tight or speed fluctuates. At the same time, when sensor data quality is unstable, time alignment errors, or calibration errors accumulate, existing systems lack branch processing paths that are tightly coupled with gripping decisions to address data quality and executability, making it difficult to obtain consistent and reproducible gripping control results under complex materials and high-speed cycle conditions.

[0005] To address this, this invention proposes an AI-based sorting machine gripping control method. This method aligns multi-source data under a unified spatial and temporal reference, generates gripping candidates through single-piece separability probability gating, and filters feasible regions based on geometric feasibility and temporal reachability. Execution candidates are selected based on gripping success probability and scoring gating. Within the vacuum establishment window, the sealing index is calculated to adaptively adjust vacuum and stripping parameters, and criterion-based branching is used to implement retry rollback and output a structured result package. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an artificial intelligence-based sorting machine gripping control method to solve the problems mentioned in the background section.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] S1. Establish a unified spatial reference and a unified time reference for the sorting machine's gripping control, and ensure that camera data, conveyor belt data, and end sensor data are time-aligned under the unified time reference;

[0009] S2. Acquire image information and form depth geometric information at the same time reference, and simultaneously acquire conveyor belt motion information and end vacuum state sequence as input for subsequent data quality gating and candidate evaluation.

[0010] S3. Obtain a set of target instances based on the image information, and output the individual separability probability of each target instance by the first artificial intelligence model; if there is a target instance that meets the gating condition, proceed to S4, otherwise output a disturbance control command and return to S2 to re-acquire.

[0011] S4. Generate grabbing candidates for target instances that meet the gating conditions, and perform geometric feasibility and temporal reachability determination on each grabbing candidate; if there is a grabbing candidate that meets both geometric feasibility and temporal reachability, proceed to S5; otherwise, output a target switching command, and output a conveyor belt adjustment command when the system allows it, and return to S2.

[0012] S5. Input the crawling candidates that simultaneously satisfy geometric feasibility and temporal reachability into the second artificial intelligence model, then output the crawling success probability, and form a scoring gating based on the crawling success probability and the single-item separability probability to select the execution candidate; when there is no execution candidate that satisfies the scoring gating, output an exception flag and return to S2.

[0013] S6. For the execution candidate, generate the end vacuum setting and stripping trajectory parameters, calculate the sealing index based on the vacuum pressure sequence and vacuum flow sequence within the vacuum establishment window, and adaptively adjust the vacuum setting and stripping trajectory parameters according to the comparison result of the sealing index and the sealing threshold; when the vacuum criterion is met and the slip criterion is not triggered, complete the trajectory planning, collision avoidance verification and capture and release, and output the structured result package; when the vacuum criterion is not met or the slip criterion is triggered, execute the release retry, and when the number of consecutive failures reaches the failure number threshold, output that the target cannot be captured and return to S2.

[0014] The further step S1, establishing a unified spatial reference, includes establishing a camera coordinate system, a robot base coordinate system, a conveyor belt coordinate system, and a sorting machine upper reference coordinate system. Establishing a unified spatial reference also includes using a calibration board to perform camera intrinsic parameter calibration and using calibration reprojection error as a verification index, using hand-eye calibration to obtain the spatial transformation relationship from the camera coordinate system to the robot base coordinate system, and obtaining the spatial transformation relationship from the conveyor belt coordinate system to the sorting machine upper reference coordinate system by identifying fixed marker points on the conveyor belt.

[0015] The establishment of a unified time reference includes generating a unified timestamp using the sorting machine controller clock and writing camera data, conveyor belt data, and end-sensor data into the unified timestamp. The camera data, conveyor belt data, and end-sensor data are then mapped to a unified sampling grid and interpolated for alignment. The calibration reprojection error threshold is 0.3 to 0.8 pixels, and the time alignment error threshold is 1 to 5 ms. When the calibration reprojection error exceeds the corresponding threshold or the time alignment error exceeds the corresponding threshold, calibration or time alignment is re-executed. If the threshold is not met after two repetitions, a calibration / synchronization failure is output, and the current round of grasping control is terminated.

[0016] The time alignment includes mapping camera frames, conveyor encoder data, and end-sensor data to a unified sampling grid and performing interpolation alignment; the time alignment error threshold is set to 1ms to 5ms, and is determined based on the upper limit of the sorting machine controller's control cycle and the statistical value of the time jitter in the acquisition link.

[0017] Furthermore, the image information in step S2 includes a color image and a depth image, the depth geometry information includes a point cloud generated from the depth image, and generating the point cloud includes performing hole repair and denoising processing on the depth image; the conveyor belt motion information includes a conveyor belt speed sequence converted from the pulse count of the conveyor belt encoder; the end vacuum state sequence includes at least a vacuum pressure sequence and a vacuum flow rate sequence, and the sampling interval between the vacuum pressure sequence and the vacuum flow rate sequence is 5ms to 20ms;

[0018] The data quality gating is determined based at least on the effective pixel ratio of depth and the speed jump amplitude of the conveyor belt. The effective pixel ratio of depth threshold is 0.85 to 0.95. The speed jump threshold is determined by collecting 500 to 2000 speed samples at rated speed during the equipment acceptance phase and taking the 95th percentile. When the data quality gating is not met, the current frame is marked as a low-quality frame and a command to wait for the next frame or a conveyor belt adjustment command is output.

[0019] Furthermore, the step S3 of obtaining the target instance set includes performing instance segmentation on the image information to output a target instance mask set; the gating condition includes a single-item separability probability not less than a separability threshold, the separability threshold being 0.60 to 0.80, and the separability threshold being selected through the validation set receiver operating characteristic curve, so that the false positive rate does not exceed 0.05 and the recall rate is not less than 0.90; the perturbation control command includes at least one of a flip command, a shake command, and a wait command, wherein the flip command or the shake command corresponds to the end executing a preset sweep trajectory, the sweep amplitude of the sweep trajectory being 5mm to 20mm and the sweep frequency being 1Hz to 3Hz.

[0020] The further step S4, generating grabbing candidates, includes sampling a set of candidate points within the target instance range that meets the gating conditions. The number of candidate points is 10 to 80, and the sampling area of ​​the candidate points is the internal area of ​​the target instance mask or the edge area of ​​the target instance mask. The geometric feasibility determination includes calculating the local flatness index and the normal angle index for the candidate points, and determining that they are geometrically feasible when the local flatness index is not greater than the flatness threshold and the normal angle index is not greater than the angle threshold. The angle threshold is 8 to 15 degrees and the flatness threshold is 0.2 square millimeters to 1.0 square millimeters. The angle threshold and the flatness threshold are determined by performing 200 to 1000 standard grabbing tests on the same material category and selecting the upper limit of the allowable grabbing success rate of not less than 0.95.

[0021] The temporal reachability determination includes calculating the predicted arrival time of the target to the reference line based on the conveyor belt speed and the target position, and estimating the time required for the robot to reach the candidate and complete the adsorption. When the sum of the current time and the required time is not greater than the predicted arrival time minus the safety margin, it is determined to be temporally reachable. The safety margin is 0.10 seconds to 0.30 seconds and is determined based on the sum of the upper bound of the controller cycle and the upper bound of the vacuum establishment time. The upper bound of the vacuum establishment time is determined by statistical analysis of no less than 300 vacuum establishment windows and taking the 99th percentile. When there is no grabbing candidate that simultaneously satisfies geometric feasibility and temporal reachability, the target switching instruction is used to abandon the current target and switch to the next target, and the conveyor belt adjustment instruction includes a speed reduction request instruction.

[0022] Furthermore, the input to the second artificial intelligence model in step S5 includes candidate point neighborhood image patches, candidate point surface feature information, and conveyor belt speed; the scoring gating is based on a combination of the success probability of grasping, the probability of single-item separability, the temporal reachability determination result, and the geometric feasibility index to obtain a score value, and the score value satisfies the following relationship:

[0023]

[0024] in To capture the success rate, For the probability of separability of a single item, c is the temporal reachability indicator, and c is the local flatness index; , , , This is the scoring weighting coefficient. The value ranges from 0.35 to 0.55. The value ranges from 0.20 to 0.35. The value ranges from 0.10 to 0.25. The scoring weight coefficient is between 0.05 and 0.15; it is determined offline through historical playback data optimization to minimize the weighted objective function of the part loss rate and the error window rate, and is then fixed as an auditable configuration.

[0025] The anomaly markers include at least one of the following: markers that require disturbance or markers that require manual intervention.

[0026] The vacuum setting in the further step S6 includes the target vacuum setting pressure. The It is a negative pressure value relative to atmospheric pressure, ranging from -35 kPa to -70 kPa;

[0027] The peeling trajectory parameters include at least the lifting height, the peeling direction angle, and the peeling speed, wherein the lifting height is 10mm to 60mm, the peeling direction angle is 0 degrees to 20 degrees, and the peeling speed is 50mm / s to 250mm / s; the vacuum establishment window duration is 0.10 seconds to 0.35 seconds; and the sealing index... The vacuum pressure sequence and vacuum flow rate sequence within the vacuum establishment window are calculated and satisfy the following relationship:

[0028]

[0029] in Where Δt is atmospheric pressure, N is the sampling interval, and N is the number of sampling points in the window. The reference flow rate is 0.2 to 0.6 times the rated flow rate of the vacuum pump.

[0030] The sealing threshold is determined according to the material category. For each material category, no less than 300 vacuum tests are collected to establish a window sequence and calculate the sealing index. The 10th percentile of the sealing index of the successfully captured sample is taken as the sealing threshold and fixed as an auditable configuration.

[0031] The adaptive adjustment includes performing at least one adjustment when the sealing index is less than the sealing threshold: increasing the magnitude of the target vacuum setting pressure, increasing the peeling direction angle, and decreasing the peeling speed; and maintaining the default vacuum setting and default peeling trajectory parameters when the sealing index is not less than the sealing threshold. The slip criterion is jointly triggered by the vacuum pressure sequence and the vacuum flow rate sequence, wherein the joint triggering satisfies that the vacuum pressure rebounds and the vacuum flow rate increases within the vacuum establishment window and lasts for 20ms to 60ms.

[0032] The threshold for consecutive failures is 2 to 5 times; the collision avoidance verification includes calculating the minimum gap between the execution trajectory and the conveyor belt, guardrail and sorting machine inlet. When the minimum gap is not less than the safety gap threshold, the grabbing and delivery are allowed. When the minimum gap is less than the safety gap threshold, the trajectory is replanned or the speed limit is reduced. The safety gap threshold is 10mm to 40mm.

[0033] The structured result package includes at least the target identifier, grasping pose, single-piece separability probability, grasping success probability, temporal reachability determination result, vacuum setting, peeling trajectory parameters, sealing index, deployment time, image evidence fragment index, and pressure-flow evidence fragment index. When grasping fails, it further includes the failure reason code, trigger threshold record, and corresponding sensor fragment index.

[0034] This invention provides an artificial intelligence-based gripping control method for sorting machines. It has the following beneficial effects:

[0035] 1. This invention constructs a gated closed loop based on the probability of single-piece separability, and triggers disturbance backoff and re-acquisition when the gate fails, so that the grasping decision is changed from "direct grasping after recognition" to a closed loop process of "entering candidate generation after separability determination", thereby reducing the failure retry caused by multiple pieces being adsorbed in one time of stacked flexible parts and misgrasping.

[0036] 2. This invention incorporates both geometric feasibility determination and temporal reachability determination into the candidate feasible domain screening, and performs branch processing such as target switching or conveyor belt adjustment when the target is unreachable, so that the candidate point selection and the sorting cycle window constraint form a consistent executable decision path, thereby reducing cycle fluctuations caused by mis-picking and temporary speed reduction.

[0037] 3. This invention calculates the sealing index based on pressure and flow sequences within the vacuum establishment window and adaptively adjusts the vacuum setting and stripping trajectory parameters accordingly. It also combines vacuum criteria and slip criteria to trigger release retry and failure rollback, transforming vacuum adsorption from single-threshold confirmation to sequence feature-driven adaptive control, thereby reducing misjudgment of part dropping and slip failure in scenarios with high permeability or easily leaky materials. Attached Figure Description

[0038] Figure 1 This is a flowchart of a sorting machine gripping control method based on artificial intelligence;

[0039] Figure 2 This is a schematic diagram illustrating the timestamp alignment of multi-source data under a unified time base according to the present invention. Detailed Implementation

[0040] To enable those skilled in the art to understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0041] The present invention will now be described in detail with reference to the accompanying drawings:

[0042] Example 1 provides an artificial intelligence-based sorting machine gripping control method applied to the suction cup peeling and gripping sorting of flexible garment parts on a conveyor belt.

[0043] This embodiment is applicable to e-commerce return sorting or warehouse shelving sorting scenarios. The items to be sorted are flexible items such as T-shirts, sweatshirts, and socks, which are folded, overlapped, and partially obscured on the conveyor belt. The end effector uses a vacuum suction cup to peel and grip the items and place them at the sorting port.

[0044] I. Equipment and Data Link Configuration

[0045] The sorting system in this embodiment includes: a sorting machine controller, a conveyor belt and its encoder, an RGB-D imaging assembly consisting of an industrial camera and a depth camera, an industrial robot and its controller, an end-effector vacuum suction cup assembly, a vacuum pressure sensor, a vacuum flow sensor, and a data acquisition board. The sorting machine controller outputs a master clock as a unified timestamp source, and the data acquisition board writes camera frames, encoder pulses, vacuum pressure data, and vacuum flow data to the same timestamp and aligns them.

[0046] II. Step S1: Establish a unified benchmark

[0047] In step S1, a unified spatial reference and a unified time reference are established.

[0048] The spatial references include the camera coordinate system, robot base coordinate system, conveyor belt coordinate system, and sorting machine upper reference coordinate system. Camera intrinsic parameters are calibrated using a calibration board, with a reprojection error threshold set to 0.5 pixels. Hand-eye calibration is used to obtain the spatial transformation from the camera coordinate system to the robot base coordinate system. Reflective dot markers are used for the fixed markers on the conveyor belt; after identification, the spatial transformation from the conveyor belt coordinate system to the sorting machine upper reference coordinate system is determined. The time reference uses the sorting machine controller clock as the master clock, with a unified timestamp written to it, and a time alignment error threshold set to 2 ms.

[0049] Verification passed: When the reprojection error is no greater than 0.5 pixels and the time alignment error is no greater than 2 ms, proceed to step S2.

[0050] Paths that fail verification: If any threshold is not met, re-execute calibration or time alignment; if it still fails after repeating twice, output "calibration / synchronization failed" and terminate this round of capture.

[0051] The thresholds are determined as follows: the reprojection error threshold is determined based on the camera resolution and on-site acceptance specifications; the time alignment error threshold is determined based on the upper bound of the controller cycle and the jitter statistics of the acquisition link.

[0052] like Figure 2As shown in the figure, the camera frame, conveyor encoder data, and end vacuum pressure and flow data are aligned under a unified timestamp. The time alignment error εt and a threshold are used for verification and branching. This demonstrates the advantages of the present invention in grasping control, which allows for verification of multi-source data time consistency and rollback of alignment anomalies. This enables subsequent target positioning, timing reachability determination, and sealing index calculation to be performed based on the same time reference, reducing window errors and misjudgments caused by time drift.

[0053] III. Step S2: Forming multimodal input

[0054] Color and depth images are acquired under a unified time reference, and point clouds are generated. At the same time, the conveyor belt speed sequence is obtained by converting encoder pulse counts, and the end vacuum pressure sequence and vacuum flow sequence are acquired.

[0055] In this embodiment, the depth map is used to generate a point cloud after hole repair and denoising; the sampling interval for vacuum pressure and vacuum flow rate is set to: Δt = 10 ms

[0056] Data quality gating is determined by the effective pixel ratio of depth and the speed jump amplitude: the effective pixel ratio threshold is set to 0.90; the speed jump threshold is determined by collecting 1000 frames of speed samples at rated speed during the equipment acceptance phase and taking the 95th percentile. In this embodiment, it is set to 0.12m / s.

[0057] Validation path: When the effective pixel ratio of depth is not less than 0.90 and the speed jump is not greater than 0.12m / s, enter S3.

[0058] Path verification failure: When the condition is not met, mark the frame as a low-quality frame and output "Wait for the next frame" or "Slowdown request", then return to S2 to re-acquire.

[0059] IV. Step S3: Separable Gated Closed Loop

[0060] Instance segmentation is performed on the target instance based on image information to obtain an instance mask set; each instance is input into the first artificial intelligence model to output the individual separability probability. The training data for the first artificial intelligence model came from stacked flexible components collected on-site, labeled as "whether a single component can be peeled off under the current contact state," and the separability threshold was selected on the validation set using ROC. This ensures that the false positive rate does not exceed 0.05 and the recall rate is not less than 0.90. In this embodiment... Take 0.70.

[0061] Path verification passed: when it exists Instances with a value ≥0.70 will enter S4.

[0062] Paths that fail validation: When all instances When all values ​​are less than 0.70, a disturbance control command is output and the system returns to S2. In this embodiment, the disturbance control command is a jitter sweeping trajectory with a sweeping amplitude of 10mm and a frequency of 2Hz, so as to cause the stacked structure to generate relative displacement and form a peelable state.

[0063] V. Step S4: Candidate Feasible Region Filtering

[0064] For a target instance that meets the gating conditions, a set of 40 candidate points is sampled within the instance mask area. Geometric feasibility indices, including local flatness *c* and the normal angle *θ*, are calculated for each candidate point. The flatness threshold and angle threshold are determined through a standard gripping test: 500 standard gripping tests are performed on the same material category, and the upper limit of the allowable gripping success rate of no less than 0.95 is selected as the threshold. In this embodiment, θ... th Take 12°, c th Take 0.6 mm².

[0065] Simultaneously perform time-series reachability determination: calculate the predicted arrival time T of the target reaching the upper reference line based on the conveyor belt speed and target position. arr And estimate the time T required for the robot to complete adsorption. rob Safety margin T safe The upper bound of the controller cycle and the upper bound of the vacuum build-up time are taken. In this embodiment, the upper bound of the controller cycle is 4ms, and the upper bound of the vacuum build-up time is 0.18s obtained by taking the 99th percentile of no less than 300 statistical analyses. Therefore, T safe Take 0.20s.

[0066] Validation path: When there is a candidate point that simultaneously satisfies both geometric feasibility and temporal reachability, proceed to S5.

[0067] Path verification fails: If the path does not exist, output the instruction "Abandon target and switch to the next target"; if the system allows it, output "Slowdown request" to the conveyor belt controller and return to S2.

[0068] VI. Step S5: Selection of Success Probability and Scoring Gating

[0069] The candidate points that satisfy both geometric and temporal feasible regions are input into the second artificial intelligence model, which outputs the probability of successful capture P. succ ; and based on P succ With P sep A scoring gating system is used to select execution candidate points. In this embodiment, the candidate scoring formula is defined as follows: A comprehensive score is given for any given candidate point:

[0070]

[0071] in: The overall score for the k-th candidate point; The probability of successfully capturing the k-th candidate point output by the second artificial intelligence model, with a value range of [0,1]. The probability of separability of a single target instance output by the first artificial intelligence model, taking the value range [0,1]. This is a time-series reachability indicator, with a value range of (0,1]. It is a geometric flatness index.

[0072] , , , These are the scoring weighting coefficients. In this embodiment, we take w1=0.45, w2=0.30, w3=0.15, and w4=0.10; the scoring threshold is Score. th The validation set is selected to ensure that the false positive rate does not exceed 0.05. In this embodiment, the score is... th Take 0.62.

[0073] Validation path: When there are candidate points with a score ≥ 0.62, select the one with the highest score to proceed to S6.

[0074] Paths that fail verification: When there are no candidate points that meet the scoring gate, an exception flag is output and S2 is returned. In this embodiment, the exception flag is "needs to be flipped".

[0075] VII. Step S6: Adaptive execution of the sealing index to close the loop and output the result package.

[0076] Set the vacuum and stripping trajectory parameters for the candidate generation process. Vacuum setting pressure. -55 kPa; the stripping trajectory parameters include the lifting height. Angle of peeling direction With peeling speed In this embodiment, =35mm =12° =160mm / s. Vacuum establishment window. The sampling time is set to 0.25s, the sampling interval Δt is 10ms, and the number of sampling points is defined as follows:

[0077]

[0078] Let the window start timestamp be T, and the m-th sampling time within the window be:

[0079]

[0080] Pressure sequence within the window With flow sequence Calculate the sealing index K sealThe K seal satisfy:

[0081]

[0082] in: Atmospheric pressure; For reference flow rate, the rated flow rate of the vacuum pump is used in this embodiment. 0.4 times.

[0083] Sealing threshold Determined by material category: At least 300 window sequences should be collected for this category and calculated. Successful sample The 10th percentile as In this embodiment: Take 0.32.

[0084] Adaptive rule: when < At that time, increase P set To -62kPa and α peel Increase to 16° and v peel Reduced to 120 mm / s; when ≥ Use the default parameters.

[0085] The slip criterion uses a combined pressure-flow trigger: triggering is determined when pressure rises and flow increases for 30ms within a window. The threshold for consecutive failures is N. fail Take 3.

[0086] Validation path: When the vacuum criterion is met and the slip criterion is not triggered, the trajectory planning and collision avoidance validation are performed to complete the capture and deployment, and the structured result package R is output.

[0087] Paths that fail verification: When the vacuum criterion is not met or the slip criterion is triggered, release and switch candidate points to retry; when three consecutive failures occur, output "Target cannot be grabbed" and return to S2.

[0088] The structured result package R should at least include: target identifier, grasp pose, , Time-series reachability determination results With peeling parameters, The system includes the delivery time, image evidence index, and pressure-flow evidence index; when a capture fails, it further includes the failure reason code, trigger threshold record, and sensor fragment index.

[0089] Example 2 provides an artificial intelligence-based sorting machine gripping control method applied to the sealing index gating and adaptive reinforcement of film packaging and highly permeable materials.

[0090] This embodiment is applicable to food or daily chemical warehousing and sorting: the items to be sorted are thin film packaging bags, breathable mesh bags, or materials whose surface texture makes the suction cups prone to leakage. The main reasons for gripping failure are "poor local sealing, unstable vacuum establishment, and high probability of slippage". This embodiment, while keeping the main process unchanged, makes targeted configurations to the sealing index threshold setting and adaptive rules in S6 to improve feasibility under highly breathable materials.

[0091] I. Key Differentiation Configurations (without changing the technical theme)

[0092] 1) Vacuum establishment window A sampling time of 0.35s and a sampling interval Δt of 5ms are used to improve the resolution of short-term leakage fluctuations; and the number of sampling points is defined as follows:

[0093]

[0094] Let the window start timestamp be T, and the m-th sampling time within the window be:

[0095] 2) Take 0.6 times the rated flow rate of the vacuum pump to increase the distinguishability of the flow rate normalization;

[0096] 3) Sealing index K seal The calculation is still performed according to the definition in Example 1:

[0097]

[0098] in: Atmospheric pressure; For reference traffic; For a moment Vacuum pressure measurement value; For a moment The vacuum flow rate measurement value.

[0099] To obtain the sealing threshold suitable for highly permeable materials, this embodiment performs offline calibration data acquisition for the same material category at two different ambient temperatures: at the first temperature setting... With the second temperature setting Under these conditions, at least 300 gripping tests were conducted; for each gripping test, the corresponding sealing index was calculated using the formula described above. .

[0100] This embodiment takes =15° =30°.

[0101] At temperature The successful sample set is defined as follows:

[0102]

[0103] The 10th percentile of each of the successful sample sets is taken as the sealing threshold for the corresponding temperature range:

[0104]

[0105]

[0106] The ambient temperature is obtained by a temperature sensor during operation. The temperature-compensated sealing threshold is obtained by linear interpolation based on the two threshold levels.

[0107]

[0108] The The sealing index is used to calculate the current capture window. A comparison is made to trigger the first or second level adaptive adjustment rules of this embodiment.

[0109] 4) The adaptive rule adopts a "two-level adjustment": Let the two-level boundary coefficient be... In this embodiment, =0.8

[0110] When satisfied Perform a first-level adjustment at the target vacuum set pressure. The downward adjustment range increased by 5 kPa, and the peeling direction angle... Increase by 4°;

[0111] The parameters corresponding to the first-level adjustment are updated as follows:

[0112]

[0113] when Perform secondary adjustment at the time, target vacuum set pressure The reduction range increased by 10 kPa, and the peeling direction angle was adjusted. Increase by 8° and peel speed Reduced by 40%;

[0114] The parameters corresponding to the level 2 adjustment are updated as follows:

[0115]

[0116] II. Example of Validation Path

[0117] In a single crawl, the first candidate point Calculated at runtime ,but If the conditions for triggering a secondary adjustment are met, the secondary adjustment will be executed. If the default parameter is... , , The parameters after the second-level adjustment are: , , Subsequently, when the vacuum criterion is satisfied and the slip criterion is not triggered, the capture and delivery are completed and a structured result package is output.

[0118] If in Internal pressure cannot be reached If the combined sliding criterion is triggered, then release and switch candidate points for retry; when the number of consecutive failures reaches [a certain threshold], [further action will be taken]. When the target is deemed ungraspable, the output is returned to S2. This embodiment, to meet high-cycle requirements, takes... .

[0119] 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 sorting machine gripping control method based on artificial intelligence, characterized in that, include: S1. Establish a unified spatial reference and a unified time reference for the sorting machine's gripping control, and ensure that camera data, conveyor belt data, and end sensor data are time-aligned under the unified time reference; S2. Acquire image information and form depth geometric information at the same time reference, and simultaneously acquire conveyor belt motion information and end vacuum state sequence as input for subsequent data quality gating and candidate evaluation. S3. Obtain a set of target instances based on the image information, and output the individual separability probability of each target instance by the first artificial intelligence model; if there is a target instance that meets the gating condition, proceed to S4, otherwise output a disturbance control command and return to S2 to re-acquire. S4. Generate grabbing candidates for target instances that meet the gating conditions, and perform geometric feasibility and temporal reachability determination on each grabbing candidate; if there is a grabbing candidate that meets both geometric feasibility and temporal reachability, proceed to S5; otherwise, output a target switching command, and output a conveyor belt adjustment command when the system allows it, and return to S2. S5. Input the crawling candidates that simultaneously satisfy geometric feasibility and temporal reachability into the second artificial intelligence model, then output the crawling success probability, and form a scoring gating based on the crawling success probability and the single-item separability probability to select the execution candidate; when there is no execution candidate that satisfies the scoring gating, output an exception flag and return to S2. S6. For the execution candidate, generate the end vacuum setting and stripping trajectory parameters, calculate the sealing index based on the vacuum pressure sequence and vacuum flow sequence within the vacuum establishment window, and adaptively adjust the vacuum setting and stripping trajectory parameters according to the comparison result of the sealing index and the sealing threshold; when the vacuum criterion is met and the slip criterion is not triggered, complete the trajectory planning, collision avoidance verification and capture and release, and output the structured result package; when the vacuum criterion is not met or the slip criterion is triggered, execute the release retry, and when the number of consecutive failures reaches the failure number threshold, output that the target cannot be captured and return to S2.

2. The method according to claim 1, characterized in that, The establishment of a unified spatial reference includes establishing a camera coordinate system, a robot base coordinate system, a conveyor belt coordinate system, and a sorting machine upper part reference coordinate system; the establishment of a unified spatial reference also includes using a calibration board to perform camera intrinsic parameter calibration and using calibration reprojection error as a verification index, using hand-eye calibration to obtain the spatial transformation relationship from the camera coordinate system to the robot base coordinate system, and obtaining the spatial transformation relationship from the conveyor belt coordinate system to the sorting machine upper part reference coordinate system by identifying fixed marker points on the conveyor belt; The establishment of a unified time reference includes generating a unified timestamp using the sorting machine controller clock and writing camera data, conveyor belt data, and end sensor data into the unified timestamp, and mapping the camera data, conveyor belt data, and end sensor data to a unified sampling grid and performing interpolation alignment. The calibration reprojection error threshold is 0.3 to 0.8 pixels, and the time alignment error threshold is 1 to 5 ms. When the calibration reprojection error exceeds the corresponding threshold or the time alignment error exceeds the corresponding threshold, the calibration or time alignment is re-executed. When the number of re-executions reaches 2 and the threshold is still not met, the calibration / synchronization failure is output and the current round of grasping control is terminated. The time alignment includes mapping camera frames, conveyor encoder data, and end-sensor data to a unified sampling grid and performing interpolation alignment; the time alignment error threshold is set to 1ms to 5ms, and is determined based on the upper limit of the sorting machine controller's control cycle and the statistical value of the time jitter in the acquisition link.

3. The method according to claim 1, characterized in that, The image information includes a color image and a depth image. The depth geometry information includes a point cloud generated from the depth image, and the point cloud generation includes hole repair and noise reduction processing on the depth image. The conveyor belt motion information includes a conveyor belt speed sequence calculated from the pulse count of the conveyor belt encoder. The end vacuum state sequence includes at least a vacuum pressure sequence and a vacuum flow rate sequence, and the sampling interval between the vacuum pressure sequence and the vacuum flow rate sequence is 5ms to 20ms. The data quality gating is determined based at least on the effective pixel ratio of depth and the speed jump amplitude of the conveyor belt. The effective pixel ratio of depth threshold is 0.85 to 0.

95. The speed jump threshold is determined by collecting 500 to 2000 speed samples at rated speed during the equipment acceptance phase and taking the 95th percentile. When the data quality gating is not met, the current frame is marked as a low-quality frame and a command to wait for the next frame or a conveyor belt adjustment command is output.

4. The method according to claim 1, characterized in that, The acquisition of the target instance set includes performing instance segmentation on the image information to output a target instance mask set; the gating condition includes a single-item separability probability not less than a separability threshold, the separability threshold being 0.60 to 0.80, and the separability threshold being selected through the validation set receiver operating characteristic curve, so that the false positive rate does not exceed 0.05 and the recall rate is not less than 0.90; the perturbation control command includes at least one of a flip command, a shake command, and a wait command, wherein the flip command or the shake command corresponds to the end executing a preset sweep trajectory, the sweep amplitude of the sweep trajectory being 5mm to 20mm and the sweep frequency being 1Hz to 3Hz.

5. The method according to claim 1, characterized in that, The generation of grabbing candidates includes sampling a set of candidate points within the target instance range that meets the gating conditions. The number of candidate points is 10 to 80, and the sampling area of ​​the candidate points is the internal area of ​​the target instance mask or the edge area of ​​the target instance mask. The geometric feasibility determination includes calculating the local flatness index and the normal angle index for the candidate points, and determining that they are geometrically feasible when the local flatness index is not greater than the flatness threshold and the normal angle index is not greater than the angle threshold. The angle threshold is 8 to 15 degrees and the flatness threshold is 0.2 square millimeters to 1.0 square millimeters. The angle threshold and the flatness threshold are determined by performing 200 to 1000 standard grabbing tests on the same material category and selecting the upper limit of the allowable grabbing success rate of not less than 0.

95. The temporal reachability determination includes calculating the predicted arrival time of the target to the reference line based on the conveyor belt speed and the target position, and estimating the time required for the robot to reach the candidate and complete the adsorption. When the sum of the current time and the required time is not greater than the predicted arrival time minus the safety margin, it is determined to be temporally reachable. The safety margin is 0.10 seconds to 0.30 seconds and is determined based on the sum of the upper bound of the controller cycle and the upper bound of the vacuum establishment time. The upper bound of the vacuum establishment time is determined by statistical analysis of no less than 300 vacuum establishment windows and taking the 99th percentile. When there is no grabbing candidate that simultaneously satisfies geometric feasibility and temporal reachability, the target switching instruction is used to abandon the current target and switch to the next target, and the conveyor belt adjustment instruction includes a speed reduction request instruction.

6. The method according to claim 1, characterized in that, The input to the second artificial intelligence model includes candidate point neighborhood image patches, candidate point surface feature information, and conveyor belt speed; the scoring gating is based on a combination of the success probability of grasping, the probability of single-item separability, the temporal reachability determination result, and the geometric feasibility index to obtain a score value, which satisfies the following relationship: in To capture the success rate, For the probability of separability of a single item, c is the temporal reachability indicator, and c is the local flatness index; , , , This is the scoring weighting coefficient. The value ranges from 0.35 to 0.

55. The value ranges from 0.20 to 0.

35. The value ranges from 0.10 to 0.

25. The scoring weight coefficient is between 0.05 and 0.15; it is determined offline through historical playback data optimization to minimize the weighted objective function of the part loss rate and the error window rate, and is then fixed as an auditable configuration. The anomaly markers include at least one of the following: markers that require disturbance or markers that require manual intervention.

7. The method according to claim 1, characterized in that, The vacuum setting in step S6 includes the target vacuum setting pressure. The It is a negative pressure value relative to atmospheric pressure, ranging from -35 kPa to -70 kPa; The peeling trajectory parameters include at least the lifting height, the peeling direction angle, and the peeling speed, wherein the lifting height is 10mm to 60mm, the peeling direction angle is 0 degrees to 20 degrees, and the peeling speed is 50mm / s to 250mm / s; the vacuum establishment window duration is 0.10 seconds to 0.35 seconds; and the sealing index... The vacuum pressure sequence and vacuum flow rate sequence within the vacuum establishment window are calculated and satisfy the following relationship: in Where Δt is atmospheric pressure, N is the sampling interval, and N is the number of sampling points in the window. The reference flow rate is 0.2 to 0.6 times the rated flow rate of the vacuum pump. The sealing threshold is determined according to the material category. For each material category, no less than 300 vacuum tests are collected to establish a window sequence and calculate the sealing index. The 10th percentile of the sealing index of the successfully captured sample is taken as the sealing threshold and fixed as an auditable configuration. The adaptive adjustment includes performing at least one adjustment when the sealing index is less than the sealing threshold: increasing the magnitude of the target vacuum setting pressure, increasing the peeling direction angle, and decreasing the peeling speed; and maintaining the default vacuum setting and default peeling trajectory parameters when the sealing index is not less than the sealing threshold. The slip criterion is jointly triggered by the vacuum pressure sequence and the vacuum flow rate sequence, wherein the joint triggering satisfies that the vacuum pressure rebounds and the vacuum flow rate increases within the vacuum establishment window and lasts for 20ms to 60ms. The threshold for consecutive failures is 2 to 5 times; the collision avoidance verification includes calculating the minimum gap between the execution trajectory and the conveyor belt, guardrail and sorting machine inlet. When the minimum gap is not less than the safety gap threshold, the grabbing and delivery are allowed. When the minimum gap is less than the safety gap threshold, the trajectory is replanned or the speed limit is reduced. The safety gap threshold is 10mm to 40mm. The structured result package includes at least the target identifier, grasping pose, single-piece separability probability, grasping success probability, temporal reachability determination result, vacuum setting, peeling trajectory parameters, sealing index, deployment time, image evidence fragment index, and pressure-flow evidence fragment index. When grasping fails, it further includes the failure reason code, trigger threshold record, and corresponding sensor fragment index.

Images