Multi-AGV-mechanical arm collaborative carrying system based on space-time constraint modeling

Abstract

The invention relates to the technical field of intelligent manufacturing, and particularly discloses a space-time constraint modeling-based multi-AGV-mechanical arm collaborative carrying system, which comprises a space-time constraint modeling module, a task decomposition and priority evaluation module, a collaborative path planning module, a dynamic collision prediction and adjustment module and an energy consumption and efficiency optimization module, according to the method, the total energy consumption, the total duration and the congestion index are subjected to normalization weighting, the unified fitness function is constructed, and the genetic algorithm and the particle swarm optimization are adopted for joint optimization, so that energy consumption minimization and operation period minimization can be taken into consideration at the same time, and a high-congestion section is actively avoided on path selection. According to the multi-target fusion, side effects caused by single index optimization are avoided, the sustainable operation capacity of the system is improved, the actual time length fed back after task execution, energy consumption and collision near-loss information are improved, model parameters are updated through a reinforcement learning or online incremental learning module, and self-evolution of a strategy is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent manufacturing technology, and in particular relates to a multi-AGV-robotic arm collaborative handling system based on spatiotemporal constraint modeling. Background Art

[0002] In today's highly automated logistics and intelligent manufacturing sectors, collaborative handling systems involving multiple AGVs (Automated Guided Vehicles) and robotic arms play a key role in improving production efficiency and reducing costs. However, existing multi-AGV-robotic arm collaborative handling systems have exposed numerous issues in practical applications, severely restricting their full potential.

[0003] Most existing systems tend to focus on a single goal when optimizing handling operations, such as simply pursuing the minimization of energy consumption or the shortest handling cycle. This single-objective optimization method has significant drawbacks. When the system focuses on reducing energy consumption, it may choose a longer handling route, resulting in a significant increase in the handling cycle, affecting the overall production progress; and when the system is committed to shortening the handling cycle, it may adopt a high-energy consumption operation method, increasing operating costs. In addition, when selecting routes, existing systems lack consideration of the degree of road congestion, which can easily cause AGVs to enter high-congestion sections, which not only increases handling time, but may also increase energy consumption due to frequent starts and stops. Due to the lack of a unified multi-objective optimization framework, existing systems find it difficult to achieve a balance between energy consumption, operation cycle and road congestion, and cannot simultaneously meet the operational requirements of high efficiency, energy saving and low congestion, resulting in limited sustainable operation capabilities of the system.

[0004] To this end, this application proposes a multi-AGV-robotic arm collaborative handling system based on spatiotemporal constraint modeling to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to provide a multi-AGV-robotic arm collaborative handling system based on spatiotemporal constraint modeling to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The multi-AGV-robotic arm collaborative handling system based on spatiotemporal constraint modeling includes:

[0008] The spatiotemporal constraint modeling module is used to collect and integrate static geometric information of the operation area, dynamic obstacle models, and traffic flow characteristics to obtain a spatiotemporal constraint model that includes operation nodes and feasible path time periods;

[0009] A task decomposition and priority assessment module is used to split the high-level handling task into several subtasks based on the spatiotemporal constraint model, and calculate the priority according to the timeliness, importance and vehicle availability of the materials to obtain a prioritized subtask list;

[0010] A collaborative path planning module is used to jointly plan the AGV travel path and the robot arm operation sequence within the spatiotemporal constraint model for each subtask in the subtask list, and obtain a path plan with a time window and a safety margin;

[0011] A dynamic collision prediction and adjustment module is used to monitor the status of the AGV and the robotic arm in real time before and during execution, and predict potential collisions based on the path plan and sensor feedback to obtain a safety redundancy adjustment plan;

[0012] The energy consumption and efficiency optimization module is used to perform multi-objective optimization evaluation on the adjustment plan, including energy consumption, transportation cycle and road congestion, to obtain the optimal execution plan.

[0013] Preferably, based on the spatiotemporal constraint model, the high-level handling task is divided into several subtasks, and the priorities are calculated according to the timeliness, importance and vehicle availability of the materials to obtain a priority-ordered subtask list, including:

[0014] Using a subtask priority calculation method, the material urgency and vehicle remaining power factors in the material information are weighted to obtain a subtask priority score;

[0015] The expression of the subtask priority calculation method is:

[0016] Si=αUi+β(1-Ei)+γWi

[0017] Where, Si: priority score of the i-th subtask;

[0018] Ui∈[0,1]: normalized value of material urgency (the larger the value, the more urgent it is);

[0019] Ei∈[0,1]: Assign the normalized value of the remaining power of the AGV (the larger the value, the more power);

[0020] Wi∈[0,1]: normalized value of material weight (the larger the value, the heavier it is);

[0021] α, β, γ: weight coefficients, satisfying α + β + γ = 1, which can be dynamically adjusted according to scenario requirements.

[0022] Preferably, for each subtask in the subtask list, the AGV travel path and the robot arm operation sequence are jointly planned within the spatiotemporal constraint model to obtain a path plan with a time window and a safety margin, including:

[0023] The LSTM-Transformer-based time series prediction method is used to jointly plan the AGV travel path and the robot arm operation timing, and a path plan with time window and safety clearance is obtained.

[0024] Preferably, the formula of the LSTM-Transformer time series prediction method is:

[0025]

[0026] in, is the predicted duration of the i-th step operation, in seconds;

[0027] hiTrans∈Rd is the hidden state vector of the Transformer decoder at step i, including historical job features and context information;

[0028] Wo∈R 1×d , bo∈R are linear mapping weights and biases, respectively, used to map the hidden vector to the duration prediction value;

[0029] d is the hidden vector dimension (e.g. 128 or 256), which is set according to the model scale.

[0030] Preferably, the state of the AGV and the robotic arm is monitored in real time before and during execution, and potential collisions are predicted based on the path plan and sensor feedback to obtain a safety redundancy adjustment plan, including: a real-time positioning unit, a collision prediction unit, and an online adjustment unit;

[0031] Wherein, the real-time positioning unit is used to obtain high-frequency position and posture data of the AGV and the robotic arm;

[0032] The collision prediction unit is configured to construct a kinematic prediction model based on the path plan and the high-frequency position and posture data, and to set a minimum safety distance threshold;

[0033] The online adjustment unit is used to automatically adjust the path or operation period to obtain the adjustment plan when the prediction model predicts that the distance is less than the minimum safety distance threshold.

[0034] Preferably, the adjustment plan is subjected to a multi-objective optimization evaluation, including energy consumption, transportation cycle and road congestion, to obtain the optimal implementation plan, including:

[0035] A multi-objective optimization algorithm combining genetic algorithm and particle swarm optimization is used to perform weighted evaluation on the total energy consumption, total distance and completion time of the adjustment plan, and output the optimal execution plan.

[0036] Preferably, the formula of the multi-objective optimization algorithm is:

[0037]

[0038] Among them, J: fitness value (the smaller the better);

[0039] Etot=∑ k PkΔtk: Total energy consumption during the entire handling process, in watt-seconds, where Pk is the power consumption during the kth segment and Δtk is the corresponding duration;

[0040] Total handling time, in seconds, is calculated by summing the predicted values ​​of the subtask priority scores in the formula;

[0041] Ccong: Traffic congestion index (such as the number of conflicts per unit time), reflecting the feasibility of the path;

[0042] Emax, Tmax, Cmax: empirical maximum values ​​of corresponding indicators, used for normalization;

[0043] w1, w2, w3: target weights, satisfying w1 + w2 + w3 = 1. They can be adjusted based on the scenario's emphasis on energy consumption or timeliness.

[0044] The total energy consumption, total transportation time and road congestion level are normalized and then weighted summed as the fitness function of genetic / particle swarm optimization, which can minimize energy consumption while ensuring high efficiency and low congestion risk.

[0045] Preferably, the system also includes a learning strategy update module, which is used to update the collaborative path planning and priority evaluation strategy parameters online after the transportation task is completed based on the actual execution log and deviation analysis report, obtain the updated strategy model, and feed it back to the spatiotemporal constraint modeling module.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] (1) This invention normalizes and weights total energy consumption, total duration, and congestion index, constructs a unified fitness function, and employs a genetic algorithm and particle swarm optimization algorithm for joint optimization. This solution simultaneously minimizes energy consumption and minimizes cycle time, while proactively avoiding highly congested sections during route selection. This multi-objective fusion avoids the side effects of optimizing a single indicator and enhances the system's sustainable operational capabilities.

[0048] (2) The present invention uses the actual duration, energy consumption, and near-miss information fed back after task execution to update model parameters via reinforcement learning or online incremental learning modules, enabling self-evolution of the strategy. As execution data continues to accumulate, the system's modeling and prediction of specific scenarios will become increasingly accurate, and the scheduling effect will be continuously optimized, with high scalability and scenario adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a block diagram of the multi-AGV-robotic arm collaborative handling system based on spatiotemporal constraint modeling of the present invention. DETAILED DESCRIPTION

[0050] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0051] Example 1:

[0052] See also Figure 1 As shown in the figure, the multi-AGV-robotic arm collaborative handling system based on spatiotemporal constraint modeling includes:

[0053] The spatiotemporal constraint modeling module is used to collect and integrate static geometric information of the operation area, dynamic obstacle models, and traffic flow characteristics to obtain a spatiotemporal constraint model that includes operation nodes and feasible path time periods;

[0054] Using a subtask priority calculation method, the material urgency and vehicle remaining power factors in the material information are weighted to obtain a subtask priority score;

[0055] The expression of the subtask priority calculation method is:

[0056] Si=αUi+β(1-Ei)+γWi

[0057] Where, Si: priority score of the i-th subtask;

[0058] Ui∈[0,1]: normalized value of material urgency (the larger the value, the more urgent it is);

[0059] Ei∈[0,1]: Assign the normalized value of the remaining power of the AGV (the larger the value, the more power);

[0060] Wi∈[0,1]: normalized value of material weight (the larger the value, the heavier it is);

[0061] α, β, γ: weight coefficients, satisfying α + β + γ = 1, which can be dynamically adjusted according to scene requirements;

[0062] By calculating the priority score for the i-th subtask, the urgency of the material's timeliness, the vehicle's remaining battery power, and the material's weight are comprehensively considered to prioritize "important and executable" tasks and reduce efficiency losses caused by insufficient energy or time delays.

[0063] A task decomposition and priority assessment module is used to split the high-level handling task into several subtasks based on the spatiotemporal constraint model, and calculate the priority according to the timeliness, importance and vehicle availability of the materials to obtain a prioritized subtask list;

[0064] A collaborative path planning module is used to jointly plan the AGV travel path and the robot arm operation sequence within the spatiotemporal constraint model for each subtask in the subtask list, and obtain a path plan with a time window and a safety margin;

[0065] A time series prediction method based on LSTM-Transformer is used to jointly plan the AGV travel path and the robot arm operation sequence, obtaining a path plan with time windows and safety clearances.

[0066] The formula of the LSTM-Transformer time series prediction method is:

[0067]

[0068] in, is the predicted duration of the i-th step operation, in seconds;

[0069] hiTrans∈Rd is the hidden state vector of the Transformer decoder at step i, including historical job features and context information;

[0070] Wo∈R 1×d , bo∈R are linear mapping weights and biases, respectively, used to map the hidden vector to the duration prediction value;

[0071] d is the hidden vector dimension (e.g. 128 or 256), set according to the model scale;

[0072] The hidden vector output by fusing the LSTM encoding sequence with the Transformer decoding context Predict the time required for step i Improve the accuracy and robustness of routing and scheduling solutions.

[0073] A dynamic collision prediction and adjustment module is used to monitor the status of the AGV and the robotic arm in real time before and during execution, and predict potential collisions based on the path plan and sensor feedback to obtain a safety redundancy adjustment plan;

[0074] The system monitors the status of the AGV and the robotic arm in real time before and during execution, and predicts potential collisions based on the path plan and sensor feedback to obtain a safety redundancy adjustment plan, including: a real-time positioning unit, a collision prediction unit, and an online adjustment unit;

[0075] Wherein, the real-time positioning unit is used to obtain high-frequency position and posture data of the AGV and the robotic arm;

[0076] The collision prediction unit is configured to construct a kinematic prediction model based on the path plan and the high-frequency position and posture data, and to set a minimum safety distance threshold;

[0077] The online adjustment unit is used to automatically adjust the path or operation period to obtain the adjustment plan when the prediction model predicts that the distance is less than the minimum safety distance threshold.

[0078] Energy consumption and efficiency optimization module, used to perform multi-objective optimization evaluation on the adjustment plan, including energy consumption, transportation cycle and road congestion, to obtain the optimal execution plan;

[0079] A multi-objective optimization algorithm combining genetic algorithm and particle swarm optimization is used to perform a weighted evaluation of the total energy consumption, total distance and completion time of the adjustment plan, and output the optimal execution plan;

[0080] The formula of the multi-objective optimization algorithm is:

[0081]

[0082] Among them, J: fitness value (the smaller the better);

[0083] Etot=∑ k PkΔtk: Total energy consumption during the entire handling process, in watt-seconds, where Pk is the power consumption during the kth segment and Δtk is the corresponding duration;

[0084] Total handling time, in seconds, is calculated by summing the predicted values ​​of the subtask priority scores in the formula;

[0085] Ccong: Traffic congestion index (such as the number of conflicts per unit time), reflecting the feasibility of the path;

[0086] Emax, Tmax, Cmax: empirical maximum values ​​of corresponding indicators, used for normalization;

[0087] w1, w2, w3: target weights, satisfying w1 + w2 + w3 = 1. They can be adjusted based on the scenario's emphasis on energy consumption or timeliness.

[0088] Normalizing the total energy consumption, total transport time, and road congestion level and then taking their weighted sum as the fitness function for genetic / particle swarm optimization can minimize energy consumption while ensuring high efficiency and low congestion risk.

[0089] The system also includes a learning strategy update module, which is used to update the collaborative path planning and priority evaluation strategy parameters online after the transportation task is completed based on the actual execution log and deviation analysis report, obtain the updated strategy model, and feed it back to the spatiotemporal constraint modeling module.

[0090] As can be seen from the above, by normalizing and weighting total energy consumption, total duration, and congestion index, constructing a unified fitness function, and jointly optimizing it with a genetic algorithm and a particle swarm algorithm, this solution can simultaneously minimize energy consumption and shorten the operation cycle, while also proactively avoiding highly congested sections during route selection. This multi-objective integration avoids the side effects of optimizing a single metric and improves the system's sustainable operational capabilities.

[0091] After task execution, actual duration, energy consumption, and near-miss information are fed back to the system, which uses reinforcement learning or online incremental learning to update model parameters and achieve self-evolution of strategies. As execution data accumulates, the system's modeling and predictions for specific scenarios become increasingly accurate, optimizing scheduling effectiveness and providing high scalability and adaptability to various scenarios.

[0092] Example 2:

[0093] Collaborative handling in small-scale sorting warehouses:

[0094] Scenario and Equipment Configuration: A 60m x 40m sorting warehouse features 30 storage nodes and two charging stations. The system deploys two AGVs (rated load 200kg, maximum speed 1.5m / s) and a robotic arm (maximum load 50kg, working radius 1.2m). Ground-based LiDAR (±2cm), UWB indoor positioning (10Hz), and a robotic arm force and torque sensor are deployed within the warehouse to obtain real-time information on work point coordinates, vehicle status, and operating torque.

[0095] Spatiotemporal Constraint Modeling: The system first uses ground-based LiDAR and UWB positioning data to automatically generate a static geometric map and mark all work nodes and charging station locations. After introducing dynamic obstacle models (such as pedestrians and forklifts), it constructs a node-time window graph, resulting in a 32-node, 240-time-interval edge structure that accurately describes the accessibility of different channels within each 0.5-second time step.

[0096] Task decomposition and priority assessment: After receiving a batch of work orders containing urgent and routine parts, the system analyzes the material timeliness, power usage rate, and weight attributes of each subtask, applies a multi-factor weighted formula (①) to calculate the priority score, and sorts the subtasks by score to ensure that tasks that can be completed with high timeliness and low remaining power are executed first.

[0097] Collaborative path and timing planning: For the first five sorted subtasks, the collaborative path planning module calls the trained LSTM-Transformer model, combines the historical operation duration sequence with the current environmental context, predicts the duration of each operation, and generates the AGV travel path and robotic arm operation period with a time window to ensure that there is no conflict between the robotic arm and AGV scheduling of the operation node.

[0098] Dynamic Collision Prediction and Online Adjustment: Before a plan is issued, the system calculates dynamic safety clearance based on the relative speed between the AGV and the end-of-arm and the set prediction window, and continuously monitors this during execution. When LiDAR or vision sensors detect a potential near-miss collision, the system automatically fine-tunes the AGV path or delays the end-of-arm operation to ensure the safety distance remains above the current threshold.

[0099] Energy consumption and efficiency multi-objective optimization: After online adjustments, the system calculates the total energy consumption, total operation time, and congestion index for each feasible solution. It then normalizes and weights each metric, using a genetic algorithm to rapidly search for the optimal solution. This optimal solution minimizes energy consumption while ensuring the shortest overall handling cycle and avoiding highly congested areas.

[0100] Learning-based strategy update: After a task is completed, the system collects actual time, energy consumption, and near-misses, and incrementally updates priority weights, duration prediction models, and safety margin parameters through a reinforcement learning algorithm, making the next round of scheduling more tailored to the actual environment.

[0101] The comparison results are shown in Table 1 below:

[0102] Table 1

[0103]

[0104]

[0105] As can be seen from the above, relying on the spatiotemporal constraint model to uniformly characterize operation nodes, access channels, and time windows, this solution can eliminate spatiotemporal conflicts during the planning phase and achieve the global optimal allocation and scheduling of transportation tasks. Compared with traditional static allocation, which only optimizes local or time-fixed paths, this solution can achieve a coordinated balance across the three dimensions of overall path, time, and resource utilization, significantly reducing idle driving and waiting costs.

[0106] By calculating subtask priorities based on multi-factor weighting (integrating urgency, battery charge, and weight), the system dynamically responds to changes in order urgency and vehicle status, ensuring that critical materials are prioritized and that tasks are not interrupted due to battery shortages. This mechanism allows for the parallel scheduling of time-sensitive tasks and stable tasks, each fulfilling its own responsibilities without interfering with the other, ensuring both efficiency and reliability.

[0107] Example 3:

[0108] Medium-sized logistics center handling and scheduling:

[0109] Scenario and device configuration:

[0110] A logistics center measuring 120m x 80m has 80 storage nodes and four charging stations. The system consists of four AGVs (300kg payload, 2.0m / s maximum speed) and a two-arm robotic arm (80kg payload, 1.5m working radius). These are equipped with multi-line LiDAR (±1cm), 20Hz UWB positioning, and a forward-looking visual obstacle avoidance camera, providing real-time feedback on scene dynamics.

[0111] Spatiotemporal constraint modeling: Through multi-source sensor fusion, a spatiotemporal graph containing 102 nodes and 640 temporal edges is generated. Traffic blockages caused by forklift movement or temporary material stacking are updated in real time to ensure that the site's accessible paths are always reflected.

[0112] Task decomposition and priority assessment: For dozens of handling tasks issued in parallel, the system combines the urgency of the materials, the power level and weight of the AGV, calculates the scores through dynamic adaptive weights (α, β, γ), and prioritizes the dispatch of urgent and high-value materials.

[0113] Collaborative Path and Timing Planning: Utilizing an enhanced dual-channel Transformer model, time series features and spatial location information are integrated in a 256-dimensional hidden layer to accurately predict the duration of the 10th step of the operation to be approximately 18 seconds. This model also generates a seamless time-windowed collaborative path for the AGV and robotic arm.

[0114] Dynamic collision prediction and online adjustment: Automatically calculates the safety clearance based on the measured relative speed between the AGV and the robotic arm and a 0.4s prediction window, and automatically increases the clearance when driving in congested areas or at high speeds to ensure zero collisions.

[0115] Energy consumption and efficiency multi-objective optimization: The total energy consumption, total duration, and congestion index are normalized and weighted, and then optimized and searched using a particle swarm optimization algorithm and genetic algorithm. Ultimately, a scheduling solution is generated that improves efficiency while minimizing energy consumption and avoiding congestion hotspots.

[0116] Learning-based strategy update: The system performs attribution analysis on duration deviations and energy consumption differences during execution, updates Transformer model weights and priority weights online, and continuously optimizes the next round of scheduling.

[0117] The comparison results are shown in Table 2 below:

[0118] Table 2

[0119]

[0120] As can be seen above, combining LSTM with a Transformer decoder to construct a deep time series prediction model can accurately predict the duration of each process by leveraging historical job features and contextual information. This prediction not only reduces timing errors in path planning but also ensures robustness to environmental perturbations through linear mapping of hidden vectors, making overall scheduling more stable and predictable.

[0121] Dynamic safety clearance calculation based on kinematic relative velocity and predicted delay enables the system to automatically adjust the safety distance threshold at different speeds and congestion levels. This strategy balances safety and efficiency—in low-risk environments, it can compress clearance to increase traffic density; in high-risk or narrow-channel environments, it can automatically increase clearance to prevent collisions.

[0122] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A multi-AGV-robotic arm collaborative handling system based on spatiotemporal constraint modeling, characterized by: include: The spatiotemporal constraint modeling module is used to collect and integrate static geometric information of the operation area, dynamic obstacle models, and traffic flow characteristics to obtain a spatiotemporal constraint model that includes operation nodes and feasible path time periods; A task decomposition and priority assessment module is used to split the high-level handling task into several subtasks based on the spatiotemporal constraint model, and calculate the priority according to the timeliness, importance and vehicle availability of the materials to obtain a prioritized subtask list; A collaborative path planning module is used to jointly plan the AGV travel path and the robot arm operation sequence within the spatiotemporal constraint model for each subtask in the subtask list, and obtain a path plan with a time window and a safety margin; A dynamic collision prediction and adjustment module is used to monitor the status of the AGV and the robotic arm in real time before and during execution, and predict potential collisions based on the path plan and sensor feedback to obtain a safety redundancy adjustment plan; The energy consumption and efficiency optimization module is used to perform multi-objective optimization evaluation on the adjustment plan, including energy consumption, transportation cycle and road congestion, to obtain the optimal execution plan.

2. The multi-AGV-robotic arm collaborative handling system based on spatiotemporal constraint modeling according to claim 1 is characterized in that: Based on the spatiotemporal constraint model, the high-level handling task is divided into several subtasks, and the priorities are calculated according to the timeliness, importance and vehicle availability of the materials to obtain a prioritized subtask list, including: Using a subtask priority calculation method, the material urgency and vehicle remaining power factors in the material information are weighted to obtain a subtask priority score; The expression of the subtask priority calculation method is: Si=αUi+β(1-Ei)+γWi Where, Si: priority score of the i-th subtask; Ui∈[0,1]: normalized value of material urgency; Ei∈[0,1]: normalized value of the remaining power of the allocated AGV; Wi∈[0,1]: normalized value of material weight; α, β, γ: weight coefficients, satisfying α + β + γ = 1.

3. The multi-AGV-robotic arm collaborative handling system based on spatiotemporal constraint modeling according to claim 1 is characterized in that: For each subtask in the subtask list, a joint planning of the AGV travel path and the robot arm operation sequence is performed within the spatiotemporal constraint model to obtain a path plan with a time window and a safety margin, including: The LSTM-Transformer-based time series prediction method is used to jointly plan the AGV travel path and the robot arm operation timing, and a path plan with time window and safety clearance is obtained.

4. The multi-AGV-robotic arm collaborative handling system based on spatiotemporal constraint modeling according to claim 3 is characterized in that: The formula of the LSTM-Transformer time series prediction method is: in, is the predicted duration of the i-th step operation, in seconds; hiTrans∈Rd is the hidden state vector of the Transformer decoder at step i, including historical job features and context information; Wo∈R 1×d , bo∈R are linear mapping weights and biases, respectively, used to map the hidden vector to the duration prediction value; d is the hidden vector dimension, which is set according to the model scale.

5. The multi-AGV-robotic arm collaborative handling system based on spatiotemporal constraint modeling according to claim 1 is characterized in that: The system monitors the status of the AGV and the robotic arm in real time before and during execution, and predicts potential collisions based on the path plan and sensor feedback to obtain a safety redundancy adjustment plan, including: a real-time positioning unit, a collision prediction unit, and an online adjustment unit; Wherein, the real-time positioning unit is used to obtain high-frequency position and posture data of the AGV and the robotic arm; The collision prediction unit is configured to construct a kinematic prediction model based on the path plan and the high-frequency position and posture data, and to set a minimum safety distance threshold; The online adjustment unit is used to automatically adjust the path or operation period to obtain the adjustment plan when the prediction model predicts that the distance is less than the minimum safety distance threshold.

6. The multi-AGV-robotic arm collaborative handling system based on spatiotemporal constraint modeling according to claim 1 is characterized in that: The adjustment plan is subjected to a multi-objective optimization evaluation, including energy consumption, transportation cycle and road congestion, to obtain the optimal execution plan, including: A multi-objective optimization algorithm combining genetic algorithm and particle swarm optimization is used to perform weighted evaluation on the total energy consumption, total distance and completion time of the adjustment plan, and output the optimal execution plan.

7. The multi-AGV-robotic arm collaborative handling system based on spatiotemporal constraint modeling according to claim 6 is characterized in that: The formula of the multi-objective optimization algorithm is: Among them, J: fitness value (the smaller the better); Etot=∑ k PkΔtk: Total energy consumption during the entire handling process, in watt-seconds, where Pk is the power consumption during the kth segment and Δtk is the corresponding duration. Total handling time, in seconds, is obtained by summing the predicted priority scores of the subtasks in the formula; Ccong: traffic congestion index, reflecting the feasibility of the route; Emax, Tmax, Cmax: empirical maximum values ​​of corresponding indicators, used for normalization; w1, w2, w3: target weights, satisfying w1 + w2 + w3 = 1, and adjusted based on the scenario's emphasis on energy consumption or timeliness.

8. The multi-AGV-robotic arm collaborative handling system based on spatiotemporal constraint modeling according to claim 1 is characterized in that: The system also includes a learning strategy update module, which is used to update the collaborative path planning and priority evaluation strategy parameters online after the transportation task is completed based on the actual execution log and deviation analysis report, obtain the updated strategy model, and feed it back to the spatiotemporal constraint modeling module.

Images