Method and system for expanding action of single-arm proxy to double-arm operation

Abstract

The invention relates to a method and a system for expanding action of a single-arm agent to double-arm operation, and belongs to the technical field of robot control. The invention designs a framework for expanding the action of a single-arm agent to double arms, the framework realizes double-arm operation by utilizing a single-arm data set, and the double-arm operation is promoted by utilizing a wide single-arm data set without additional double-arm data. According to the scheme, an attention expansion mechanism is introduced, semantic information is embedded into pixel-level data through the mechanism, and action pairs are allowed to be dynamically generated. These actions process the proposed TAN (Target Allocation Network) that allocates tasks between two robotic arms, assesses feasibility, and selects an optimal solution. In addition, according to the scheme, seamless transition to single-arm operation can be achieved when abnormity is detected, and it is ensured that tasks are executed continuously. Experimental results show that the technical scheme provided by the invention improves the task efficiency and eliminates the cost of collecting special double-arm data at the same time.

Description

Technical Field

[0001] The present invention belongs to the technical field of robot control, and relates to a method and a system for extending the action of a single-arm agent to a dual-arm operation. Background Art

[0002] In recent years, the application of dual-arm robots in many fields has achieved remarkable results. They not only play the role of intelligent assistants in the home environment, such as performing daily housework and caring for the elderly and children, but also show unparalleled precision and collaboration on industrial assembly lines, greatly improving production efficiency and product quality. At the same time, in the agricultural field, dual-arm robots also provide strong support for the planting, harvesting and subsequent processing of crops with their powerful operating capabilities and adaptability, helping to promote agricultural modernization. Behind these successes is the result of the continuous advancement and innovation of dual-arm robot technology, especially its outstanding performance in flexibility, coordination and efficient execution of complex tasks.

[0003] However, as the demand for dual-arm robots in various fields continues to grow, a significant problem has gradually surfaced: How to obtain high-quality dual-arm datasets to effectively train dual-arm manipulation strategies? This challenge mainly stems from the complexity and diversity of dual-arm interactions. In practical applications, dual-arm robots need to deal with target coordination, object attribute differences, and various uncertain factors in the real world, such as changes in ambient lighting and object position offsets, which greatly increase the difficulty and cost of data collection. Traditional data collection methods are often time-consuming and labor-intensive, and it is difficult to fully cover all possible interaction scenarios, resulting in limited generalization capabilities of trained models.

[0004] In addition, most current dual-arm manipulation strategies are based on predefined rigid rules and lack sufficient flexibility and adaptability to cope with real-time changing situations. For example, during the execution of a task, if an arm fails or encounters unexpected task irregularities, existing strategies are often difficult to make timely adjustments, thus affecting the smooth completion of the overall task. Therefore, it is particularly important to develop a more dynamic and responsive dual-arm manipulation strategy.

[0005] In this context, extending the existing single-arm agent algorithms to the field of dual-arm manipulation has become a highly promising solution. The advantage of this approach is that it can make full use of the large number of existing single-arm datasets, and through algorithm expansion and transfer learning, it can achieve effective training of dual-arm agents, thereby significantly reducing the cost and time of data collection. However, this expansion process is not easy. Existing single-arm agent algorithms are mainly designed for sequential, single-target tasks. When faced with scenarios that require simultaneous manipulation of multiple objects or the performance of complex coordination tasks, their efficiency will be greatly reduced. This is mainly because single-arm algorithms lack the ability to handle complex mechanisms such as dual-target action generation, allocation, their interdependencies, and spatial relationship coordination.

[0006] Therefore, to truly achieve efficient training and operation of dual-arm intelligent bodies, it is necessary to conduct in-depth innovation at the algorithm level and develop new algorithms that can effectively manage the generation and allocation of dual-target actions and achieve efficient coordination and synchronization between arms. Through such expansion and optimization, not only can the existing single-arm dataset resources be fully utilized, but also the execution efficiency and adaptability of dual-arm robots in various complex tasks can be significantly improved, laying a solid foundation for the widespread application of dual-arm robot technology. Summary of the invention

[0007] In view of this, the purpose of the present invention is to provide a method and system for extending the actions of a single-arm agent to dual-arm operations. The method and system provide a framework called ExS2D (Extending the Actions of a Single-Arm Agent to Dual Arms), which includes two major components: 1) a target generation network; 2) a target allocation network. The target generation network embeds semantic information into pixel-level data by introducing an attention expansion mechanism, deeply mines and fully utilizes the features of semantic data and visual data, and dynamically generates action pairs. The target allocation network allocates the action pairs dynamically generated by the target generation network to the two robotic arms according to the state of the current environment, evaluates the feasibility and selects the optimal solution to achieve efficient dual-arm operation. In particular, when a robotic arm is detected to be abnormal, the action pair originally performed by the robotic arm will be automatically transferred to another robotic arm to ensure uninterrupted task execution.

[0008] In order to achieve the above object, the present invention provides the following technical solutions:

[0009] A method for extending the actions of a single-arm agent to a two-arm operation, the method specifically comprising the following steps:

[0010] S1: Build an attention expansion module that uses global visual information o and natural language instructions l to get the first picking point;

[0011] S2: Build an object segmentation module to generate a mask for the first object based on the first picking point.

[0012] S3: Based on the first object mask, the attention expansion module is used again to generate the second picking point;

[0013] S4: construct a transport module, which uses global visual information, natural language instructions, and the coordinates of the first pick-up point to obtain the coordinates of the first placement point;

[0014] S5: Based on the coordinates of the first placement point, the object segmentation module is used again to generate a mask for the first placement area;

[0015] S6: Based on the first placement area mask, the transport module is used again to generate the second placement point;

[0016] S7: construct a state encoder that uses the target state graph to obtain the encoding features of the target state graph;

[0017] S8: construct a cost encoder that uses the cost matrix to obtain encoding features of the cost matrix;

[0018] S9: Construct a target assignment module, which uses the encoding features output by the state encoder and the encoding features output by the cost encoder, and combines the motion planning library to achieve effective coordination, so as to assign the target action to the corresponding robotic arm.

[0019] Furthermore, in step S1, an attention extension module is constructed, which obtains a probability map of the best picking point of the first object by fusing the natural language instructions with the visual data. The pixel with the highest probability value on the probability map is the best picking point of the object. The coordinates of the first picking point are obtained by the following formula:

[0020]

[0021] Among them, Q pick is the action value function for the picking operation, and (u,v) is the pixel coordinate on the image.

[0022] Further, in step S2, an object segmentation module is constructed. Based on the generated coordinates of the first picking point, a 64×64 pixel cropping block is cropped with the point as the center. The cropping block can completely contain the first object. Then, the mask of the first object is obtained by using SAM, and the global image containing the mask of the first object is converted into m1 As new visual information for use in subsequent processes.

[0023] Further, in step S3, the natural language instruction l and the global image o obtained in step S2 are used. m1, the attention expansion module is used again to obtain the probability map of the best picking point for the second object. Since the first object has been processed by SAM, the attention expansion module can focus on the second object, thereby obtaining the second picking point more accurately; similar to solving the coordinates of the first picking point, the pixel with the highest probability in the probability map is the best picking point. The coordinates of the second picking point are obtained by the following formula:

[0024]

[0025] In step S4, a transport module is constructed. The input visual information and natural language instructions are consistent with those input in step S1. The transport module uses these data and combines the coordinates of the first pickup point to generate a probability map for the first placement point. The pixel with the largest probability value in the probability map is the best placement point for the first object. The coordinates of the first placement point are obtained by the following formula:

[0026]

[0027] Among them, Q place is the action value function for the placement operation, and Δτ is the pixel coordinate on the image.

[0028] Further, in step S5, a SAM module is constructed to crop a 8×8 pixel crop block centered on the first placement point, and a mask is generated for the area using SAM to ensure that the transport module focuses on generating the possible placement area of ​​the second object; the global image o containing the mask of the first placement point area is m2 It is used as new visual information for subsequent solution of the second placement point.

[0029] Further, in step S6, the natural language instruction l and the global image o obtained in step S5 are combined. m2 And the coordinates of the second picking point obtained in step S3, the transport module is used again to obtain the probability map of the second placement point. The pixel with the maximum probability value in the probability map is the best placement point for the second object; the coordinates of the second placement point are obtained by the following formula:

[0030]

[0031] Further, in step S7, a state encoder is constructed, which captures the spatial relationship between the target and the dual arms by processing the observation results in the target state graph, and embeds the position data of the task and the manipulator into the feature vector using MLP and MHSA; the target node in the target state graph contains the action posture of the corresponding object in the Cartesian coordinate system, that is, the pick-up and placement coordinates of the object, which are expressed by the following formula:

[0032]

[0033] T1 and T2 represent the coordinates of picking and placing respectively; the warehouse node in the target state diagram contains the starting position of the corresponding robot arm, which is expressed by the following formula:

[0034]

[0035] Arm1 and Arm2 represent the Cartesian coordinates of the two robotic arms respectively.

[0036] Further, in step S8, a cost encoder is constructed, which collects the execution costs from the cost matrix. As the robot performs the pick and place actions, the cost matrix changes dynamically with the changing working state; the cost matrix consists of a 2×2 grid, and each grid cell contains two costs: the movement cost c mv and transfer costs c tf , where the moving cost represents the distance the robot moves from the current position to the target object; the transfer cost represents the distance the target object moves from the pick-up position to the placement position; the two costs are solved using the following formula:

[0037]

[0038] In the case of infeasible actions, such as assigning the same task to both arms or assigning an already completed task, the cost in the respective grid cell is set to infinity. The cost matrix captures the path lengths and corresponding transfer costs from the robot's starting position to the two action pairs:

[0039]

[0040] Among them, cost(T i ,Arm j ) is the total path length of the robot arm from its starting position to the picking position and then completing the delivery; when a robot arm encounters a failure, the action that should have been performed by this robot arm will be seamlessly transferred to another robot arm to ensure uninterrupted execution of the action.

[0041] Further, in step S9, a task allocation module is constructed, which uses the MHSA layer to fuse the cooperation information of the target node and the cost matrix, and then applies the MLP layer and the probability maximization operation to output the optimal task allocation strategy for both arms; the entire learning process optimizes the allocation of actions by minimizing the cross entropy loss function;

[0042] The final output of the target allocation network is generated based on the cost matrix, where each task is assigned to the robot arm with the lowest execution cost. The output of the target allocation network is expressed as:

[0043]

[0044] Among them, A(1) and A(2) indicate that the two action pairs are assigned to the robot arm respectively. j and Arm k .

[0045] The present invention also provides a system for extending the actions of a single-arm agent to a dual-arm operation.

[0046] The beneficial effects of the present invention are:

[0047] The present invention provides a framework called ExS2D (Extending the Action of Single-Arm Agent to Dual-Arm), which includes two major components: 1) target generation network; 2) target allocation network. The target generation network embeds semantic information into pixel-level data by introducing an attention extension mechanism, deeply mines and fully utilizes the features of semantic data and visual data to dynamically generate action pairs. The target allocation network allocates the action pairs dynamically generated by the target generation network to the two robotic arms according to the state of the current environment, evaluates the feasibility and selects the optimal solution to achieve efficient dual-arm operation. In particular, when a robotic arm is detected to be abnormal, the action pair originally performed by the robotic arm will be automatically transferred to the other robotic arm to ensure uninterrupted task execution.

[0048] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0050] Figure 1 is a system block diagram of the present invention;

[0051] Figure 2 Illustrative example diagrams of target generation modules;

[0052] Figure 3 Illustrative example diagram of assigning modules to targets. DETAILED DESCRIPTION

[0053] The technical solution of the present invention is described in detail below with reference to the accompanying drawings.

[0054] The framework provided by the present invention for extending the actions of a single-arm agent to dual-arm operations is used to promote dual-arm operations by utilizing a data set of single-arm operations. The present invention proposes a new framework, including the implementation of the following nine stages: 1) constructing an attention extension module, using visual information and natural language instructions to obtain the coordinates of the first picking point; 2) generating a mask of the corresponding object around the first picking point through the object segmentation module; 3) using the attention extension module again, obtaining the coordinates of the second picking point according to the visual data and natural language instructions containing the mask of the first picking object; 4) in the transport module, combining the initial visual data, natural language instructions and the coordinates of the first picking point to obtain the coordinates of the first placement point; 5) generating a mask of the placement area around the first placement point through the object segmentation module; 6) using the transport module again, obtaining the coordinates of the second placement point according to the visual data, natural language instructions and the coordinates of the second picking point containing the mask of the first placement area; 7) constructing a state encoder, using the target state diagram to obtain the encoding features of the target state diagram; 8) constructing a cost encoder, using the cost matrix to obtain the encoding features of the cost matrix; 9) constructing a target allocation module, which uses the encoding features output by the state encoder and the encoding features output by the cost encoder, and combines the motion planning library to achieve effective coordination, thereby allocating the target action to the corresponding robot arm. This framework utilizes a large number of easily accessible single-arm operation datasets to promote dual-arm operation, thereby solving the problem of difficulty in obtaining dual-arm operation datasets. Figure 1 It is a system block diagram of the present invention.

[0055] In this embodiment, four necessary concepts are involved, which are as follows:

[0056] The first concept: global visual information, which is a top-down RGB-D three-dimensional reconstructed image captured by three noise-free cameras. The resolution of each camera is 640×480. Two of the cameras are located on the left and right sides of the table, and the remaining camera is located directly in front of the robotic arm. All three cameras are facing the center of the table.

[0057] The second concept: natural language instructions. Natural language instructions are instructions given by humans in the expectation that the robotic arm can complete the tasks mentioned therein.

[0058] The third concept: target state graph, the target state graph is constructed as a non-fully connected graph that reflects the spatial relationship and task completion status in the current workspace. It consists of several target nodes and 2 warehouse nodes. The target node represents the pick and place action, and each node contains the action posture in the absolute Cartesian coordinates of the object. The warehouse node represents the starting point of the two robotic arms.

[0059] The fourth concept: cost matrix, which represents the cooperation cost of all potential actions in the current workspace state. These costs are affected by the target-related information and the configuration of the two arms during the execution of the action.

[0060] Figure 2 Illustrative example diagram of the target generated network, Figure 3 This is an illustrative example diagram of the target allocation network. The system framework of this system mainly includes two components: target generation network and target allocation network. Figure 1 The system block diagram of the present invention is shown, wherein: Target generation network: The starting point of the network is to use global visual information and natural language instructions to obtain the optimal pick-up point coordinates of the target object, including the following steps:

[0061] Step 1: Input global visual information and natural language instructions, and obtain the coordinates of the first object picking point through the attention expansion module;

[0062] Step 2: Use the object segmentation module to obtain the mask of the first object;

[0063] Step 3: Based on the global visual information and natural language instructions including the mask of the first object, the attention expansion module is used again to obtain the coordinates of the second object picking point;

[0064] Step 4: Based on the global visual information, natural language instructions and the coordinates of the first pick-up point, the transport module is used to obtain the coordinates of the first placement point;

[0065] Step 5: Around the first placement point, use the object segmentation module again to generate a mask for the first placement point area;

[0066] Step 6: Based on the global visual information including the first object placement area mask, natural language instructions, and the second pickup point coordinates, the transport module is used again to obtain the second placement point coordinates.

[0067] Goal Assignment Network: The starting point of this network is to assign the action pairs generated by the Goal Generation Network to the two robotic arms with the lowest execution cost. It includes the following steps:

[0068] Step 1: Input the target state diagram into the state encoder to obtain the encoding features of the target state diagram;

[0069] Step 2: Input the cost matrix into the cost encoder to obtain the encoding features of the cost matrix;

[0070] Step 3: Fusion the encoding features output by the state encoder and the encoding features output by the cost encoder, and use the target assignment module to assign the target action pairs to the two robotic arms respectively.

[0071] In this embodiment, an attention extension module is constructed, which obtains a probability map of the best picking point of the first object by fusing natural language instructions with visual data. The pixel with the highest probability value on the probability map is the best picking point of the object. The coordinates of the first picking point are obtained by the following formula:

[0072]

[0073] Among them, Q pick is the action value function for the picking operation, and (u,v) is the pixel coordinate on the image.

[0074] The method also includes: constructing an object segmentation module, based on the generated coordinates of the first picking point, cutting out a 64×64 pixel cropping block with the point as the center; the cropping block can completely contain the first object, and then using SAM to obtain the mask of the first object, and converting the global image containing the mask of the first object into m1 As new visual information for use in subsequent processes.

[0075] Use the natural language instruction l and the obtained global image o m1 , the attention expansion module is used again to obtain the probability map of the best picking point for the second object. Since the first object has been processed by SAM, the attention expansion module can focus on the second object, thereby obtaining the second picking point more accurately; similar to solving the coordinates of the first picking point, the pixel with the highest probability in the probability map is the best picking point. The coordinates of the second picking point are obtained by the following formula:

[0076]

[0077] Build a transport module. The transport module uses this data and the coordinates of the first pickup point to generate a probability map for the first placement point. The pixel with the largest probability value in the probability map is the best placement point for the first object. The coordinates of the first placement point are obtained by the following formula:

[0078]

[0079] Among them, Q placeis the action value function for the placement operation, and Δτ is the pixel coordinate on the image.

[0080] Construct a SAM module, crop an 8×8 pixel crop block centered on the first placement point, and use SAM to generate a mask for the area to ensure that the transport module focuses on generating the possible placement area of ​​the second object; m2 It is used as new visual information for subsequent solution of the second placement point.

[0081] In this embodiment, the natural language instruction l and the global image o are combined. m2 And the coordinates of the second pickup point, use the transport module again to obtain the probability map of the second placement point. The pixel with the largest probability value in the probability map is the best placement point for the second object; the coordinates of the second placement point are obtained by the following formula:

[0082]

[0083] A state encoder is constructed to capture the spatial relationship between the target and the dual arms by processing the observations in the target state graph. The position data of the task and the manipulator are embedded into the feature vector using MLP and MHSA. The target node in the target state graph contains the action posture of the corresponding object in the Cartesian coordinate system, that is, the pick-up and placement coordinates of the object, which are expressed by the following formula:

[0084]

[0085] T1 and T2 represent the coordinates of picking and placing respectively; the warehouse node in the target state diagram contains the starting position of the corresponding robot arm, which is expressed by the following formula:

[0086]

[0087] Arm1 and Arm2 represent the Cartesian coordinates of the two robotic arms respectively.

[0088] Construct a cost encoder that collects execution costs from a cost matrix. As the robot performs pick and place actions, the cost matrix changes dynamically with the changing working state; the cost matrix consists of a 2×2 grid, and each grid cell contains two costs: the movement cost c mv and transfer costs c tf , where the moving cost represents the distance the robot moves from the current position to the target object; the transfer cost represents the distance the target object moves from the pick-up position to the placement position; the two costs are solved using the following formula:

[0089]

[0090] In the case of infeasible actions, such as assigning the same task to both arms or assigning an already completed task, the cost in the respective grid cell is set to infinity. The cost matrix captures the path lengths and corresponding transfer costs from the robot's starting position to the two action pairs:

[0091]

[0092] Among them, cost(T i ,Arm j ) is the total path length of the robot arm from its starting position to the picking position and then completing the delivery; when a robot arm encounters a failure, the action that should have been performed by this robot arm will be seamlessly transferred to another robot arm to ensure uninterrupted execution of the action.

[0093] Construct a task allocation module that uses the MHSA layer to fuse the collaborative information of the target node and the cost matrix, and then applies the MLP layer and the probability maximization operation to output the optimal task allocation strategy for both arms; the entire learning process optimizes the allocation of actions by minimizing the cross entropy loss function;

[0094] The final output of the target allocation network is generated based on the cost matrix, where each task is assigned to the robot arm with the lowest execution cost. The output of the target allocation network is expressed as:

[0095]

[0096] Among them, A(1) and A(2) indicate that the two action pairs are assigned to the robot arm respectively. j and Arm k .

[0097] Experimental results show that the technical solution provided by the present invention improves task efficiency while eliminating the cost of collecting specialized dual-arm data.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A method for extending the actions of a single-arm agent to a two-arm operation, characterized in that: The method specifically comprises the following steps: S1: Build an attention expansion module that uses global visual information o and natural language instructions l to get the first picking point; S2: Build an object segmentation module to generate a mask for the first object based on the first picking point. S3: Based on the first object mask, the attention expansion module is used again to generate the second picking point; S4: construct a transport module, which uses global visual information, natural language instructions, and the coordinates of the first pick-up point to obtain the coordinates of the first placement point; S5: Based on the coordinates of the first placement point, the object segmentation module is used again to generate a mask for the first placement area; S6: Based on the first placement area mask, the transport module is used again to generate the second placement point; S7: construct a state encoder that uses the target state graph to obtain the encoding features of the target state graph; S8: construct a cost encoder that uses the cost matrix to obtain encoding features of the cost matrix; S9: Construct a target assignment module, which uses the encoding features output by the state encoder and the encoding features output by the cost encoder, and combines the motion planning library to achieve effective coordination, so as to assign the target action to the corresponding robotic arm.

2. A method for extending the action of a single-arm agent to a dual-arm operation according to claim 1, characterized in that: In step S1, an attention extension module is constructed, which obtains a probability map of the best picking point of the first object by fusing natural language instructions with visual data. The pixel with the highest probability value on the probability map is the best picking point of the object. The coordinates of the first picking point are obtained by the following formula: Among them, Q pick is the action value function for the picking operation, and (u,v) is the pixel coordinate on the image.

3. A method for extending the action of a single-arm agent to a dual-arm operation according to claim 2, characterized in that: In step S2, an object segmentation module is constructed. Based on the generated coordinates of the first picked point, a 64×64 pixel cropping block is cropped with the point as the center. The cropping block can completely contain the first object. Then, the SAM is used to obtain the mask of the first object. The global image containing the mask of the first object is o m1 As new visual information for use in subsequent processes.

4. A method for extending the action of a single-arm agent to a dual-arm operation according to claim 3, characterized in that: In step S3, the natural language instruction l and the global image o obtained in step S2 are used. m1 , the attention expansion module is used again to obtain the probability map of the best picking point for the second object. Since the first object has been processed by SAM, the attention expansion module can focus on the second object, thereby obtaining the second picking point more accurately; similar to solving the coordinates of the first picking point, the pixel with the highest probability in the probability map is the best picking point. The coordinates of the second picking point are obtained by the following formula: In step S4, a transport module is constructed. The input visual information and natural language instructions are consistent with those input in step S1. The transport module uses these data and combines the coordinates of the first pickup point to generate a probability map for the first placement point. The pixel with the largest probability value in the probability map is the best placement point for the first object. The coordinates of the first placement point are obtained by the following formula: Among them, Q place is the action value function for the placement operation, and Δτ is the pixel coordinate on the image.

5. A method for extending the action of a single-arm agent to a dual-arm operation according to claim 4, characterized in that: In step S5, a SAM module is constructed to crop a 8×8 pixel cropped block centered on the first placement point, and a mask is generated for the area using SAM to ensure that the transport module focuses on generating the possible placement area of ​​the second object. The global image o containing the mask of the first placement point area is m2 It is used as new visual information for subsequent solution of the second placement point.

6. A method for extending the actions of a single-arm agent to a dual-arm operation according to claim 5, characterized in that: In step S6, the natural language instruction l and the global image o obtained in step S5 are combined. m2 And the coordinates of the second picking point obtained in step S3, the transport module is used again to obtain the probability map of the second placement point. The pixel with the maximum probability value in the probability map is the best placement point for the second object; the coordinates of the second placement point are obtained by the following formula:

7. A method for extending the action of a single-arm agent to a dual-arm operation according to claim 6, characterized in that: In step S7, a state encoder is constructed to capture the spatial relationship between the target and the two arms by processing the observations in the target state graph, and the position data of the task and the manipulator are embedded into the feature vector using MLP and MHSA; The target node in the target state diagram contains the action posture of the corresponding object in the Cartesian coordinate system, that is, the pick-up and placement coordinates of the object, which is expressed by the following formula: T1 and T2 represent the coordinates of picking and placing respectively; the warehouse node in the target state diagram contains the starting position of the corresponding robot arm, which is expressed by the following formula: Arm1 and Arm2 represent the Cartesian coordinates of the two robotic arms respectively.

8. A method for extending the action of a single-arm agent to a dual-arm operation according to claim 7, characterized in that: In step S8, a cost encoder is constructed to collect execution costs from a cost matrix. As the robot performs pick-up and place actions, the cost matrix changes dynamically with the changing working state. The cost matrix consists of a 2×2 grid, and each grid cell contains two costs: the movement cost c mv and transfer costs c tf , where the moving cost represents the distance the robot moves from the current position to the target object; the transfer cost represents the distance the target object moves from the pick-up position to the placement position; the two costs are solved using the following formula: c mv =||Arm k -T i pk ||,c tf =||T i pk -T i pl The cost matrix captures the path lengths from the robot’s starting position to two action pairs and the corresponding transfer costs: Among them, cost(T i ,Arm j ) is the total path length of the robot arm from its starting position to the picking position and then completing the delivery; when a robot arm encounters a failure, the action that should have been performed by this robot arm will be seamlessly transferred to another robot arm to ensure uninterrupted execution of the action.

9. A method for extending the actions of a single-arm agent to a dual-arm operation according to claim 8, characterized in that: In step S9, a task allocation module is constructed, which uses the MHSA layer to fuse the cooperation information of the target node and the cost matrix, and then applies the MLP layer and the probability maximization operation to output the optimal task allocation strategy for both arms; The entire learning process optimizes the distribution of actions by minimizing the cross entropy loss function; The final output of the target allocation network is generated based on the cost matrix, where each task is assigned to the robot arm with the lowest execution cost. The output of the target allocation network is expressed as: {A(1),A(2)}={(T1 pk ,T1 pl )→Arm j ,(T2 pk ,T2 pl )→Arm k } Among them, A(1) and A(2) indicate that the two action pairs are assigned to the robot arm respectively. j and Arm k .

10. A system for extending the actions of a single-arm agent to a dual-arm operation, characterized in that: The system adopts the method as claimed in any one of claims 1 to 9.

Images