Method and system for controlling motion trail stability when mechanical arm grabs large thin plate

Abstract

The invention discloses a motion track stability control method and system when a mechanical arm grabs a large thin plate, and the method comprises the steps: collecting environment data around the mechanical arm and the thin plate in real time through a camera and a sensor, and creating an operation environment model; initializing a motion path of the mechanical arm based on the working environment model, representing the motion path as a graph structure, inputting the graph structure into a pre-trained graph neural network model, and planning the motion path of the mechanical arm through the pre-trained graph neural network model to obtain a planned path under the current working environment model; and a control task of the mechanical arm is executed through cooperation of the multiple agents, control decision optimization is conducted by adopting a multi-agent reinforcement learning strategy, and the mechanical arm is made to move according to the planned path. The method can quickly adapt to the dynamically changing working environment, adjust the planned path in real time, improve the reliability of path planning, and improve the stability of track control of the mechanical arm on the basis of guaranteeing the grabbing precision.

Description

Technical Field

[0001] The invention belongs to the technical field of robot control, and in particular relates to a method and system for smoothly controlling the motion trajectory of a robot arm when it grabs a large thin plate. Background Art

[0002] With the rapid development of modern automation technology, robots, as key execution equipment, are widely used in various fields such as manufacturing, assembly, warehousing, and logistics. The grasping motion of the robot arm is one of its important functions, especially when handling objects with complex shapes and different weights. The smoothness of its motion trajectory directly affects the efficiency and accuracy of grasping. For example, in the process of grasping and handling large thin plate materials by a robot arm, since thin plate materials usually have a large surface area and light weight, they are more susceptible to factors such as airflow, vibration, and changes in the center of gravity under high-speed motion, which makes the trajectory control during the grasping process particularly complex and challenging.

[0003] Although traditional robot control methods can achieve object grasping to a certain extent, they often experience trajectory jitter, deviation or grasping failure when facing dynamic environments and uncertain factors. They cannot meet the requirements for grasping accuracy and stability during high-speed movement, which may cause the thin plate to shift, deform or be damaged. In addition, during high-speed grasping movements, the inertia and external interference of the robot arm will also lead to insufficient stability, further affecting the safety and reliability of the operation. With the development of artificial intelligence and deep learning technology, the use of intelligent control algorithms to optimize the real-time trajectory of the robot arm has become a research hotspot, providing new opportunities for more intelligent and efficient grasping control. However, it is still difficult to achieve efficient and accurate trajectory planning and control in the face of complex and dynamic working environments.

[0004] Therefore, an efficient method for smooth control of the grasping motion trajectory is needed to improve the adaptability of the robotic arm in complex working environments and to achieve smooth trajectory control of the robotic arm when grasping large thin plates and moving at high speed. Summary of the invention

[0005] In view of this, the present invention proposes a method and system for smoothly controlling the motion trajectory of a robotic arm when grasping a large thin plate, which is used to solve the problem that the robotic arm is difficult to achieve accurate trajectory planning and control when facing a complex and dynamic working environment.

[0006] In a first aspect, the present invention discloses a method for smoothly controlling the motion trajectory of a robot arm when grasping a large thin plate, the method comprising:

[0007] The camera and sensors are used to collect environmental data around the robot arm and the thin plate in real time to create a working environment model;

[0008] Initialize the motion path of the robot arm based on the working environment model, represent the motion path as a graph structure and input it into a pre-trained graph neural network model, plan the motion path of the robot arm through the pre-trained graph neural network model, and obtain the planned path under the current working environment model;

[0009] The control task of the robotic arm is performed collaboratively by multiple intelligent agents, and the control decision is optimized by using a multi-agent reinforcement learning strategy, so that the robotic arm moves according to the planned path.

[0010] On the basis of the above technical solution, preferably, the motion path of the robot arm is initialized based on the working environment model, the motion path is represented as a graph structure and input into a pre-trained graph neural network model, and the motion path of the robot arm is planned by the pre-trained graph neural network model, and the planned path under the current working environment model is obtained, which specifically includes:

[0011] Extract key path nodes in the working environment model and generate initial motion trajectories through a hybrid reinforcement learning strategy;

[0012] Representing the initial motion trajectory as a graph structure, inputting the graph structure into a pre-trained graph neural network model, and outputting a planned path under the current working environment model;

[0013] When the working environment model changes, the planned path is represented as a graph structure, and the motion path of the robot arm is replanned through a pre-trained graph neural network model to update the planned path under the current working environment model.

[0014] On the basis of the above technical solutions, preferably, the hybrid reinforcement learning strategy is obtained by weighted fusion of a value-based reinforcement learning algorithm and a strategy-based reinforcement learning algorithm.

[0015] On the basis of the above technical solution, preferably, in the process of planning the motion path of the robotic arm through the pre-trained graph neural network model, the node update calculation formula of the graph structure is as follows:

[0016] h p (l+1)=AGGREGATE({h q (l) , e pq |q∈N(p)})

[0017] Among them, h q (l) represents the feature vector of node q in the kth network layer, h p (l+1) represents the feature vector of node p in the l+1th network layer, e pqrepresents the edge feature between node p and its neighbor node q, N(p) represents the set of neighbor nodes of node p; AGGREGATE(·) is the AGGREGATE aggregation function.

[0018] On the basis of the above technical solution, preferably, the decision objective function of the control decision optimization using the multi-agent reinforcement learning strategy is:

[0019]

[0020] Among them, J(Π1,Π2,…,Π i ,…,Π n ) is the decision objective function, i = 1, 2, ..., n, n is the total number of agents, Π i represents the decision of the ith agent, r∈(0,1) is the discount factor used to weigh the importance of future rewards; t=0,1,…,T, T is the preset decision cycle, R i (S t ,A t,i ) is the reward function of the ith agent at time step t, S t is the state at time step t, A t,i is the action of the ith agent at time step t.

[0021] On the basis of the above technical solution, preferably, the method further comprises:

[0022] Use distributed computing architecture to optimize the motion path of the robot arm;

[0023] Each computing node of the distributed computing architecture independently executes a deep reinforcement learning algorithm to optimize the motion path of the robot arm, and the objective function of each computing node is:

[0024]

[0025] Among them, J k (θ) is the objective function of the kth computational node, represents the strategy defined by the shared parameter θ on the kth computing node, τ is a target motion trajectory, r(s t ,a t ) is the state of the robot arm at time t t Execute action a t The corresponding rewards; Indicated in strategy The expectation of the total reward obtained along the target trajectory τ;

[0026] The distributed computing architecture updates the shared parameter θ through the global coordinator, integrates the motion path optimization results of each computing node, and obtains the optimized planning path; the formula for updating the shared parameter θ is:

[0027]

[0028] Among them, θ′ is the updated shared parameter, α is the learning rate, which is used to control the step size of the update. is the objective function J k (θ) is the gradient of the shared parameters θ.

[0029] On the basis of the above technical solution, preferably, the method further comprises:

[0030] The quantum annealing algorithm is used to solve the combinatorial optimization problem in motion path optimization.

[0031] In a second aspect, the present invention discloses a motion trajectory smooth control system for a robot arm grasping a large thin plate, the system comprising:

[0032] Environmental acquisition module: used to collect environmental data around the robot arm and thin plate in real time through cameras and sensors to create an operating environment model;

[0033] Path planning module: used to initialize the motion path of the robot arm based on the working environment model, represent the motion path as a graph structure and input it into a pre-trained graph neural network model, plan the motion path of the robot arm through the pre-trained graph neural network model, and obtain the planned path under the current working environment model;

[0034] Trajectory control module: used to execute the control task of the robot arm through multiple agents in collaboration, and use multi-agent reinforcement learning strategy to optimize the control decision, so that the robot arm moves according to the planned path. In a third aspect of the present invention, an electronic device is disclosed, including: at least one processor, at least one memory, a communication interface and a bus;

[0035] Wherein, the processor, memory, and communication interface communicate with each other via the bus;

[0036] The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to implement the method according to the first aspect of the present invention.

[0037] According to a fourth aspect of the present invention, a computer-readable storage medium is disclosed, wherein the computer-readable storage medium stores computer instructions, and the computer instructions enable a computer to implement the method described in the first aspect of the present invention.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1) The present invention collects environmental data around the robot arm and the thin plate in real time to create an operating environment model; plans the motion path of the robot arm based on the operating environment model and a pre-trained graph neural network model to obtain the planned path under the current operating environment model. The present invention can quickly adapt to the dynamically changing operating environment, adjust the planned path in real time, and improve the reliability of path planning; the present invention adopts a multi-agent reinforcement learning strategy to optimize control decisions, which can improve the stability of the robot arm trajectory control.

[0040] 2) When the working environment model changes, the present invention represents the planned path as a graph structure, re-plans the motion path of the robot arm through a pre-trained graph neural network model, and updates the planned path under the current working environment model by updating the nodes of the graph structure. It can adapt to the dynamic environment and realize efficient and accurate trajectory planning in a complex and dynamic working environment.

[0041] 3) The present invention distributes the control task of the robotic arm to multiple intelligent agents, each of which controls a part of the robotic arm. By taking advantage of information sharing between intelligent agents, the present invention adopts a multi-agent reinforcement learning strategy to optimize the control decision, and designs an objective function for the control decision optimization to achieve smooth control of the overall trajectory. On the basis of ensuring the accuracy of trajectory control, the stability of large-scale thin plate grabbing operations is improved.

[0042] 4) The present invention adopts a distributed computing architecture to optimize the motion path of the robot arm, and uses a quantum annealing algorithm to solve the combinatorial optimization problem in path planning, thereby improving the overall computing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0044] Figure 1 The present invention is a flow chart of a method for smoothly controlling the motion trajectory of a robot arm when grasping a large thin plate. DETAILED DESCRIPTION

[0045] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] See also Figure 1 The present invention discloses a method for smoothly controlling the motion trajectory of a robot arm when grabbing a large thin plate, the method comprising:

[0047] S1. Collect environmental data around the robot arm and thin plate in real time through cameras and sensors to create an operating environment model.

[0048] In an embodiment of the present invention, a camera, distance sensor, pressure sensor, etc. installed on a robotic arm are used to collect environmental data of the working area in real time, the collected environmental data is preprocessed and fused, and a three-dimensional working environment model is created.

[0049] Specifically, the image data collected by the camera is preprocessed by noise filtering, feature enhancement and feature extraction; the data collected by the sensor is noise filtered and normalized to remove noise and outliers. Then, the data from the camera and sensor are feature matched and fused, and the working environment is 3D modeled through 3D reconstruction algorithms such as structured light and stereo vision to generate a 3D working environment model.

[0050] S2. Initialize the motion path of the robot arm based on the working environment model, represent the motion path as a graph structure and input it into a pre-trained graph neural network model, plan the motion path of the robot arm through the pre-trained graph neural network model, and obtain the planned path under the current working environment model.

[0051] Step S2 specifically includes the following sub-steps:

[0052] S21. Extract the key path nodes in the working environment model and generate the initial motion trajectory through a hybrid reinforcement learning strategy.

[0053] In an embodiment of the present invention, the hybrid reinforcement learning strategy is obtained by weighted fusion of a value-based reinforcement learning algorithm (such as Q-learning) and a policy-based reinforcement learning algorithm (such as a policy gradient method).

[0054] The hybrid reinforcement learning strategy of the present invention is described below by taking Q-learning and the policy gradient method as examples.

[0055] According to the working environment model, the key path nodes are identified to form a node diagram. The key path nodes are usually turning points in the grasping process, obstacle locations, edges of thin plates, etc.

[0056] The grasping task is modeled as a reinforcement learning environment, and a reinforcement learning framework is designed, including state space, action space, and reward mechanism. The state space is defined as the position of the robot arm, the current environment information, and the relative position to the target object (thin plate). The action space includes various motion strategies that the robot arm may adopt, such as movement direction, speed, acceleration and other parameters. The reward mechanism refers to the design of a reward function, which gives positive rewards when the robot arm approaches the critical path node and completes the grasping task, and negative rewards when it produces too much vibration and fails to grasp successfully.

[0057] The Q-learning and policy gradient methods are weighted and integrated to form a hybrid reinforcement learning strategy. In each training step, when selecting an action based on the current state, the value of all possible actions is first evaluated through Q-learning, and then the weighted strategy output probability is used to select an action. The calculation formula of the hybrid reinforcement learning strategy is:

[0058] Π H (a|s)=ω1Π V (a|s)+ω2Π AS (a|s)

[0059] Among them, H (a|s) is the expression of the mixed reinforcement learning strategy, Π V (a|s) is the value-based reinforcement learning algorithm, Π AS (a|s) is a policy-based reinforcement learning algorithm, ω1 and ω2 are fusion weights, ω1+ω2=1, a is the action of the robot, and s is the current state of the robot.

[0060] Conduct multiple rounds of training in a defined reinforcement learning environment, and continuously adjust model parameters to improve policy accuracy. Generate the initial motion trajectory using the action sequence output by the hybrid strategy, and further optimize the generated initial motion trajectory based on feedback information and simulation results, adjust the movement speed, and optimize the path.

[0061] S22. Represent the initial motion trajectory as a graph structure, input the graph structure into a pre-trained graph neural network model, and output the planned path under the current working environment model.

[0062] The present invention represents the motion trajectory of the robotic arm as a graph structure, which represents the overall structural information in the form of a graph, including the connection relationship between nodes and edges, where nodes represent key points on the path, and point features may include position, speed, direction, etc. Edges represent the relationship between paths, and edge features may be distance, interaction force, etc.

[0063] In an embodiment of the present invention, the robot arm grasping operation in different working environments is simulated by simulation software, the motion trajectory of the robot arm is converted into a graph structure, different training graphs are generated, and an optimal trajectory is planned for each grasping operation. The graph structure corresponding to the optimal trajectory is used as the label of the training graph to construct a diversified data set. The graph neural network model is trained by the data set, and the trained graph neural network model can learn the trajectory optimization rules under different working environment conditions and perform trajectory planning.

[0064] The initial motion trajectory is represented as a graph structure, and the graph structure is input into a pre-trained graph neural network model to output the planned path under the current working environment model.

[0065] S23. When the working environment model changes, the planned path is represented as a graph structure, the motion path of the robot arm is replanned through a pre-trained graph neural network model, and the planned path under the current working environment model is updated.

[0066] In the embodiment of the present invention, each node p is defined to have a feature vector h p (l) represents the state of the node in the lth network layer. In the process of planning the motion path of the robot arm through the pre-trained graph neural network model, the feature vector of each node p is updated from its neighbor node q through the AGGREGATE aggregation function in the l+1 layer. The node update calculation formula of the graph structure is as follows:

[0067] h p (l+1)=AGGREGATE({h q (l) , e pq |q∈N(p)})

[0068] Among them, h q (l) represents the feature vector of node q in the kth network layer, h p (l+1) represents the feature vector of node p in the l+1th network layer, e pq represents the edge feature between node p and its neighbor node q, N(p) represents the set of neighbor nodes of node p; AGGREGATE(·) is the aggregation function in the graph neural network model.

[0069] The present invention represents the planned path as a graph structure. When the working environment model changes, the motion path of the robot arm is re-planned through a pre-trained graph neural network model, and the planned path under the current working environment model is updated by updating the nodes of the graph structure. It can adapt to dynamic environments and achieve efficient and accurate trajectory planning in complex and dynamic working environments.

[0070] S3. Use distributed computing architecture to optimize the motion path of the robot arm.

[0071] In the embodiment of the present invention, the path optimization task is assigned to multiple computing nodes by designing and deploying a distributed computing architecture. Each computing node of the distributed computing architecture independently executes a deep reinforcement learning algorithm to optimize the local motion path, and then updates the shared parameters through a global coordinator, integrates the optimization results of each node, and obtains the optimized planning path to improve the overall computing efficiency and trajectory stability.

[0072] In the distributed computing architecture of the present invention, the objective function of each computing node is:

[0073]

[0074] Among them, J k (θ) is the objective function of the kth computing node, θ is a shared parameter, represents the strategy determined by the shared parameter θ on the kth computing node, representing an optimized motion path; τ is a target motion trajectory, r(s t ,a t ) is the state of the robot arm at time t t Execute action a t The corresponding reward function; Indicated in strategy The expectation of the total reward obtained along the target trajectory τ.

[0075] The distributed computing architecture of the present invention updates the shared parameter θ through a global coordinator to integrate the motion path optimization results of each computing node. The formula for updating the shared parameter θ is:

[0076]

[0077] Among them, θ′ is the updated shared parameter, α is the learning rate, which is used to control the step size of the update. is the objective function J k (θ) is the gradient of the shared parameters θ.

[0078] S4. Use quantum annealing algorithm to solve combinatorial optimization problems in motion path optimization.

[0079] The combinatorial optimization problem in path planning mainly involves finding the optimal path or optimal movement strategy under given constraints, which usually includes comprehensive consideration of multiple objective functions and constraints.

[0080] In an embodiment of the present invention, a quantum annealing algorithm is used to process combinatorial optimization problems in path planning, giving full play to the parallel processing capability of quantum computing to achieve more efficient trajectory stabilization control.

[0081] The objective function of path optimization using the quantum annealing algorithm is:

[0082]

[0083] Among them, H is the total path cost, C uv is the cost of selecting the path where nodes u and v are located, is the state of path node u. When path node u is selected, the value is 1, otherwise it is 0; is the state of path node v. When path node v is selected, the value is 1, otherwise it is 0.

[0084] The present invention aims to minimize the total path cost H and solves the optimal path from multiple alternative planning paths.

[0085] The present invention adopts a distributed computing architecture to optimize the motion path of the robot arm, and uses a quantum annealing algorithm to solve the combinatorial optimization problem in path planning, which can improve the overall computing efficiency, quickly adapt to changes in the working environment, and perform path optimization in real time.

[0086] S5. The control task of the robotic arm is performed collaboratively by multiple intelligent agents, and the control decision is optimized by using the Multi-Agent Reinforcement Learning (MARL) strategy, so that the robotic arm moves according to the planned path.

[0087] In an embodiment of the present invention, multiple agents are installed on the robot arm, and the control task of the robot arm is assigned to the multiple agents, each of which controls a part of the robot arm. The information sharing and coordination decision criteria between the agents are defined, and the control task of the robot arm is executed by multiple agents in a coordinated manner, and the control decision is optimized by using a multi-agent reinforcement learning strategy to achieve smooth control of the overall trajectory.

[0088] The decision objective function of the control decision optimization using multi-agent reinforcement learning strategy is:

[0089]

[0090] Among them, J(Π1,Π2,…,Π i ,…,Π n ) is the decision objective function, i = 1, 2, ..., n, n is the total number of agents, Π i represents the decision of the ith agent, r∈(0,1) is the discount factor used to weigh the importance of future rewards; t=0,1,…,T, T is the preset decision cycle, R i (S t ,A t,i) is the i-th agent performing action A at time step t t,i The reward function, S t is the state of the robot at time step t, A t,i is the action of the ith agent at time step t.

[0091] When a robot grabs a large thin plate, it is generally necessary to meet the requirements of grabbing accuracy and speed constraints. For the grabbing operation of a large thin plate, the smoothness of the motion trajectory must also be guaranteed, such as reducing the vibration or tilt of the large thin plate during the movement to prevent the thin plate from offsetting, colliding or deforming. Therefore, designing a reasonable reward function is crucial to optimizing the control decision of multiple agents in the task of grabbing large thin plates by a robot. In order to ensure that each agent can work together and achieve a smooth motion trajectory in the task of grabbing large thin plates by a robot, an embodiment of the present invention designs a comprehensive reward function scheme based on the considerations of grabbing accuracy, motion smoothness, speed constraints and thin plate stability.

[0092] In the embodiment of the present invention, the reward function R i (S t ,A t,i ) is:

[0093] R i (S t ,A t,i )=R i_pos +R i_smooth +R i_speed +R i_stability

[0094] Among them, R i_pos is the grasping accuracy reward of the i-th agent, which is used to encourage the end effector of the robot arm to reach the target position accurately.

[0095]

[0096] Among them, ε is the grasping accuracy weight factor, are the target position of the robot arm and the current position of the end effector, respectively.

[0097] R i_smooth is the motion smoothness reward for the ith agent, which aims to promote smooth motion by limiting the change in acceleration and avoiding vibrations and sudden movements.

[0098]

[0099] Among them, β is the grasping accuracy weight factor, acc i is the acceleration of the robot controlled by the ith agent at time step t. var(acc i) is the variance of historical acceleration, reflecting the stability of motion; acc max is the predefined maximum acceptable acceleration.

[0100] R i_speed is the speed reward of the ith agent, which is used to ensure that the robot arm works at an appropriate speed.

[0101]

[0102] Among them, V i is the current velocity of the robot controlled by the ith agent, V max is the maximum allowed speed, γ is the speed reward weight; δ is the speeding penalty factor.

[0103] R i_stability is the stability reward of the ith agent, which is used to ensure the stability of the large thin plate during movement. The present invention measures the smoothness of the motion trajectory of the robot arm when grabbing the large thin plate by monitoring the inclination angle of the thin plate.

[0104] R i_stability =λ·(1-tilt angle )

[0105] Among them, tilt angle is the tilt angle of the plate, ranging from 0 to 1, where 0 means complete stability; λ is the stability weight, with higher rewards for smaller tilt angles.

[0106] According to the specific task requirements of the robot arm in the process of grasping large thin plates, the present invention takes into account the grasping accuracy, movement smoothness, speed constraint and thin plate stability, designs a reward function for each agent, and adopts a multi-agent reinforcement learning strategy to optimize the control decision, which can effectively improve the trajectory control accuracy and the smoothness of the grasping motion.

[0107] The present invention proposes a method for controlling the smooth motion trajectory of a robot arm when grasping a large thin plate. By real-time collection of environmental data around the robot arm and the thin plate, an operating environment model is created. The motion path of the robot arm is planned based on the operating environment model and a pre-trained graph neural network model. The planned path under the current operating environment model is obtained, which can quickly adapt to the dynamically changing operating environment, adjust the planned path in real time, and improve the reliability of path planning. In addition, the present invention adopts a multi-agent reinforcement learning strategy to optimize control decisions, thereby improving the stability of the robot arm trajectory control while ensuring the grasping accuracy.

[0108] On the basis of the above method embodiments, the present invention further discloses a motion trajectory smoothing control system when a robot arm grabs a large thin plate, the system comprising:

[0109] Environmental acquisition module: used to collect environmental data around the robot arm and thin plate in real time through cameras and sensors to create an operating environment model;

[0110] Path planning module: used to initialize the motion path of the robot arm based on the working environment model, represent the motion path as a graph structure and input it into a pre-trained graph neural network model, plan the motion path of the robot arm through the pre-trained graph neural network model, and obtain the planned path under the current working environment model;

[0111] Trajectory control module: It is used to execute the control task of the robotic arm through the collaboration of multiple intelligent agents, and adopt the multi-agent reinforcement learning strategy to optimize the control decision so that the robotic arm moves according to the planned path.

[0112] The above system embodiments and method embodiments correspond one to one, and for a brief description of the system embodiments, please refer to the method embodiments.

[0113] The present invention also discloses an electronic device, comprising: at least one processor, at least one memory, a communication interface and a bus; wherein the processor, memory, and communication interface communicate with each other through the bus; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to implement the aforementioned method of the present invention.

[0114] The present invention also discloses a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, wherein the computer instructions enable the computer to implement all or part of the steps of the method described in the embodiment of the present invention. The storage medium includes: a USB flash drive, a mobile hard disk, a read-only memory ROM, a random access memory RAM, a magnetic disk or an optical disk, and other media that can store program codes.

[0115] The system embodiment described above is merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be distributed to multiple network units. A person skilled in the art may select some or all of the modules according to actual needs to achieve the purpose of the solution of this embodiment without creative effort.

[0116] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for smoothly controlling the motion trajectory of a robot arm when grabbing a large thin plate, characterized in that: The method comprises: The camera and sensors are used to collect environmental data around the robot arm and the thin plate in real time to create a working environment model; Initialize the motion path of the robot arm based on the working environment model, represent the motion path as a graph structure and input it into a pre-trained graph neural network model, plan the motion path of the robot arm through the pre-trained graph neural network model, and obtain the planned path under the current working environment model; The control task of the robotic arm is performed collaboratively by multiple intelligent agents, and the control decision is optimized by using a multi-agent reinforcement learning strategy, so that the robotic arm moves according to the planned path.

2. The method for controlling the motion trajectory of a robot arm when grabbing a large thin plate according to claim 1 is characterized in that: Initializing the motion path of the robot arm based on the working environment model, representing the motion path as a graph structure and inputting the pre-trained graph neural network model, planning the motion path of the robot arm through the pre-trained graph neural network model, and obtaining the planned path under the current working environment model specifically include: Extract key path nodes in the working environment model and generate initial motion trajectories through a hybrid reinforcement learning strategy; Representing the initial motion trajectory as a graph structure, inputting the graph structure into a pre-trained graph neural network model, and outputting a planned path under the current working environment model; When the working environment model changes, the planned path is represented as a graph structure, and the motion path of the robot arm is replanned through a pre-trained graph neural network model to update the planned path under the current working environment model.

3. The method for controlling the stable motion trajectory of a robot arm when grabbing a large thin plate according to claim 2 is characterized in that: The hybrid reinforcement learning strategy is obtained by weighted fusion of a value-based reinforcement learning algorithm and a strategy-based reinforcement learning algorithm.

4. The method for controlling the motion trajectory of a robot arm when grabbing a large thin plate according to claim 1 is characterized in that: In the process of planning the motion path of the robotic arm through the pre-trained graph neural network model, the node update calculation formula of the graph structure is as follows: h p (l+1)=AGGREGATE({h q (l) , e pq |q∈N(p)}) Among them, h q (l) represents the feature vector of node q in the kth network layer, h p (l+1) represents the feature vector of node p in the l+1th network layer, e pq represents the edge feature between node p and its neighbor node q, N(p) represents the set of neighbor nodes of node p; AGGREGATE(·) is the AGGREGATE aggregation function.

5. The method for controlling the stable motion trajectory of a robot arm when grabbing a large thin plate according to claim 1 is characterized in that: The decision objective function of the control decision optimization using multi-agent reinforcement learning strategy is: Among them, J(Π1,Π2,…,Π i ,…,Π n ) is the decision objective function, i = 1, 2, ..., n, n is the total number of agents, Π i represents the decision of the ith agent; r∈(0,1) is the discount factor used to weigh the importance of future rewards; t=0,1,…,T, T is the preset decision cycle, R i (S t ,A t,i ) is the reward function of the ith agent at time step t, S t is the state at time step t, A t,i is the action of the ith agent at time step t.

6. The method for controlling the stable motion trajectory of a robot arm when grabbing a large thin plate according to claim 1, characterized in that: The method further comprises: Use distributed computing architecture to optimize the motion path of the robot arm; Each computing node of the distributed computing architecture independently executes a deep reinforcement learning algorithm to optimize the motion path of the robot arm, and the objective function of each computing node is: Among them, J k (θ) is the objective function of the kth computation node, represents the strategy defined by the shared parameter θ on the kth computing node, τ is a target motion trajectory, r(s t ,a t ) is the state of the robot arm at time t t Execute action a t The corresponding rewards; Indicated in strategy The expectation of the total reward obtained along the target trajectory τ; The distributed computing architecture updates the shared parameter θ through the global coordinator, integrates the motion path optimization results of each computing node, and obtains the optimized planning path; the formula for updating the shared parameter θ is: Among them, θ′ is the updated shared parameter, α is the learning rate, is the objective function J k (θ) is the gradient of the shared parameters θ.

7. The method for controlling the stable motion trajectory of a robot arm when grabbing a large thin plate according to claim 6 is characterized in that: The method further comprises: The quantum annealing algorithm is used to solve the combinatorial optimization problem in motion path optimization.

8. A control system for the stable motion trajectory of a robot arm when grabbing a large thin plate, characterized in that: The system comprises: Environmental acquisition module: used to collect environmental data around the robot arm and thin plate in real time through cameras and sensors to create an operating environment model; Path planning module: used to initialize the motion path of the robot arm based on the working environment model, represent the motion path as a graph structure and input it into a pre-trained graph neural network model, plan the motion path of the robot arm through the pre-trained graph neural network model, and obtain the planned path under the current working environment model; Trajectory control module: It is used to execute the control task of the robotic arm through the collaboration of multiple intelligent agents, and adopt the multi-agent reinforcement learning strategy to optimize the control decision so that the robotic arm moves according to the planned path.

9. An electronic device, characterized in that: include: at least one processor, at least one memory, a communication interface, and a bus; Wherein, the processor, memory, and communication interface communicate with each other via the bus; The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions enable a computer to implement the method according to any one of claims 1 to 7.

Images