Robot autonomous assembly method and system based on imitation learning, storage medium and computer device
By combining imitation learning, multi-task learning, and meta-learning algorithms, a general expert policy and parallel encoding attention network are constructed, which solves the problem of insufficient generalization ability of imitation learning in polygonal shaft hole assembly, and realizes rapid learning and generalization of robot autonomous assembly.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTH CHINA UNIV OF TECH
- Filing Date
- 2025-05-28
- Publication Date
- 2026-07-31
AI Technical Summary
Existing imitation learning methods have insufficient generalization ability in unseen environments or situations, and struggle to learn and generalize quickly when the training data distribution is significantly different, resulting in negative transfer problems.
By combining imitation learning, multi-task learning, and meta-learning algorithms, a general expert policy applicable to polygonal shaft-hole assembly is constructed. A database is built through demonstration data collection, and a parallel encoding attention network and loss function are used to guide the algorithm to train the general assembly model, achieving rapid generalization.
It solves the problems of time-consuming, labor-intensive, and inconsistent demonstration data collection, enables rapid learning and generalization, improves the efficiency and accuracy of robot autonomous assembly, and reduces learning costs.
Smart Images
Figure CN120663302B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to robotics technology, specifically to a method, system, storage medium, and computer equipment for autonomous robot assembly based on imitation learning. Background Technology
[0002] The essence of robot assembly skill learning is learning expert strategies for adjusting assembly actions based on perceived force / torque. Currently, policy learning methods are mainly divided into two categories: one is policy learning methods based on reinforcement learning frameworks. These methods can achieve autonomous learning through interaction with various assembly environments without prior knowledge. The other is policy learning methods based on supervised learning frameworks. These methods summarize human assembly strategies from human instruction. Compared to reinforcement learning algorithms, imitation learning algorithms do not require interaction with the environment, avoiding the exploration and trial-and-error process, and thus can learn and imitate human strategies more quickly. However, imitation learning models often have insufficient generalization ability in unseen environments or situations. Performance degradation occurs in new tasks.
[0003] Research shows that meta-learning can improve a model's rapid generalization ability on tasks with scarce data through cross-task training. Model-Agnostic Meta-Learning (MAML), as a state-of-the-art representative meta-learning algorithm, is widely used in various fields. The key idea of MAML is to learn the optimal initial model parameters through cross-task training, helping the model adapt quickly to new tasks with only a small number of gradient updates. MAML's task independence allows it to be flexibly applied to any model learned through gradient descent. However, MAML suffers from the problems of requiring a large number of similar meta-tasks for training and slow convergence. Furthermore, for tasks with significantly different training data distributions, MAML struggles to learn effective shallow features, which affects the performance of meta-learning.
[0004] To address the aforementioned issues, meta-transfer learning algorithms are researched. These algorithms utilize transfer learning to transfer pre-trained models on multi-task datasets to meta-learning algorithms, thereby improving the model's convergence speed and generalization performance. Research also explores combining MAML with transfer learning algorithms to enhance the model's generalization ability. While these methods can accelerate algorithm convergence, transfer learning suffers from negative transfer problems. Achieving rapid skill learning and generalization with limited data remains a challenge. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a robot autonomous assembly method based on imitation learning. This robot autonomous assembly method based on imitation learning solves the problems of time-consuming and labor-intensive demonstration data collection and difficulties in skill learning caused by inconsistent demonstrations.
[0006] The second objective of this invention is to provide a robot autonomous assembly system based on imitation learning.
[0007] A third objective of this invention is to provide a storage medium.
[0008] A fourth objective of this invention is to provide a computer device.
[0009] The first objective of this invention is achieved through the following technical solution: This robot autonomous assembly method based on imitation learning includes the following steps:
[0010] S1. Analyze and summarize the expert experience in assembling polygonal shafts and holes, and construct a general expert strategy for assembling polygonal shafts and holes;
[0011] S2. Based on a general expert strategy, data is collected through demonstration to build a database;
[0012] S3. Construct a general assembly model based on shared knowledge and task-specific knowledge:
[0013] [ W v, F ω] T =p·[ W f- W f d , F τ- F τ d ] T +ζ,
[0014] Where p and ζ represent skill control parameters, which are related to the assembly task; W f d , F τ d Thresholds representing the contact states of each assembly;
[0015] S4. Based on the database, a loss function is used to guide the algorithm training and optimization of the general assembly model;
[0016] S5. Deploy the optimized general assembly model into the robot's control system to drive the robot to perform human-like assembly actions.
[0017] Step S2 includes the following steps:
[0018] The S21 six-degree-of-freedom industrial robot has a camera attached to its end effector to capture task-type images and collect real-time images of the expert's hand movements. At the same time, during the expert demonstration, a six-dimensional force sensor is used to measure the force and torque of the expert's hand contact during the demonstration.
[0019] S22. The images captured by the camera are transmitted to the controller via USB to calculate the position and posture of the hands during the expert demonstration;
[0020] S23. The force sensor's measured value is converted into a voltage signal by a converter, processed by the Beckhoff module and converted into a digital quantity. This digital quantity and the encoder values of each joint of the robot are transmitted to the controller via the EtherCAT protocol.
[0021] Step S3 includes the following steps:
[0022] S31. Construct a Gaussian mixture regression model to learn shared knowledge from the data collected from expert demonstrations;
[0023] S32. Construct a parallel encoding attention network. This parallel encoding attention network is based on a database and outputs the learned task-specific knowledge.
[0024] S33. Construct a multi-task meta-learning generalization framework to guide and encode attention network training in parallel, enabling rapid generalization of knowledge for specific tasks.
[0025] S34. Construct a general assembly model based on the obtained shared knowledge and the generalized specific tasks.
[0026] The parallel encoding attention network mentioned in step S32 includes:
[0027] The task-level encoder employs an encoder-decoder model architecture, integrating multi-layer 2D convolutional networks to extract implicit task-level information from acquired task-type images.
[0028]
[0029] in, It is the input task scene image; This indicates the extracted implicit task information; E task (·) and D task (·) represent the task encoder function and the task decoder function, respectively; This indicates the task scenario for refactoring;
[0030] The action layer encoder adopts an encoder-decoder model architecture, integrating multiple temporal convolutional networks to extract implicit action layer information from the acquired assembly force / torque signals:
[0031]
[0032] in, It is an input force / torque signal segment; This indicates the extracted implicit action information; E action(·) and D action (·) represents the motion encoder function and the motion decoder function; This represents the reconstructed force / torque signal segment;
[0033] The attention aggregation strategy employs an attention mechanism to analyze the importance of information at different levels. It weights and aggregates information by calculating a normalized importance matrix to achieve task-specific learning.
[0034]
[0035] Where W1, W2, and W3 represent three learnable parameter matrices, and FC(·) represents a fully connected layer.
[0036] Step S33 includes the following steps:
[0037] S331. Multi-task learning is employed to jointly train all tasks in multi-task learning to learn the optimal parallel coding applicable to all assembly tasks. Note the network parameters; the mathematical expression for parameter updates in multi-task learning is:
[0038]
[0039] Where, θ MTL γ represents the parameters of a parallel encoded attention network trained using multi-task learning; MTL Indicates the update step size for multi-task learning; N represents the loss value for the i-th training task; task The number of tasks in the training task set; Represents task η i Supporting dataset;
[0040] S332. Meta-learning cross-task training is used to learn a set of general parallel encoding attention network parameters, which are then transferred to the new assembly task through gradient updates.
[0041] S333. Combining multi-task learning and meta-learning algorithms to achieve rapid generalization of parameters to specific task knowledge, the update formula for the combined multi-task learning-meta-learning algorithm is as follows:
[0042]
[0043] or,
[0044] Multi-task learning and meta-learning are combined using a mutual supervision mechanism to achieve rapid generalization of knowledge for specific tasks based on parameters. The expression for this is:
[0045]
[0046] Wherein, θMTL θ represents the parameters of a parallel encoded attention network trained using multi-task learning; Meta The basic HIL parameters for meta-learning training; γ MTL Indicates the update step size for multi-task learning; N represents the loss value for the i-th training task; task The number of tasks in the training task set; Represents task η i Supporting datasets; EE * The expected number of training iterations during parameter backtracking is represented by EE; EE represents the current number of training iterations.
[0047] Step S332, the cross-task training process of meta-learning, includes inner loop update and outer loop update. The expression for the inner loop update is:
[0048]
[0049] in, It is based on task η i Updated parallel encoding attention network parameters; θ Meta For general parallel encoding attention network parameters trained by meta-learning; α Meta This indicates the step size for updating the inner loop; Represents task η i Supporting dataset;
[0050] The expression for updating the outer loop is:
[0051]
[0052] Where, β Meta Update the step size for the outer loop; Represents task η i The query dataset.
[0053] The loss function guiding the algorithm in step S4 includes the reconstruction loss function and the action error loss function.
[0054] The reconstruction error loss function is used as an optimization metric to minimize the error between the reconstructed task scenario and force / torque signal and the input task scenario and force / torque signal. Its mathematical expression is:
[0055]
[0056] in, This indicates the task scenario for refactoring; It is the input task scene image; This represents the reconstructed force / torque signal segment; It is an input force / torque signal segment;
[0057] The action error loss function, which aims to minimize the error between the predicted expert action and the actual expert action, is mathematically expressed as follows:
[0058]
[0059] in, Indicates the actions of prediction experts; W v k , F ω k ] T This indicates the actions of a real expert.
[0060] The second objective of this invention is achieved through the following technical solution: a robot autonomous assembly system based on imitation learning, used to implement the robot autonomous assembly method based on imitation learning described in the first objective, comprising:
[0061] The strategy formulation module is used to analyze and summarize the experience of experts in polygonal shaft and hole assembly, and to construct a general expert strategy applicable to polygonal shaft and hole assembly.
[0062] The data acquisition and processing module is used to collect expert demonstration data and establish a database for training.
[0063] The model module is used to build a general assembly model based on general knowledge and task-specific knowledge, and to train and optimize the general assembly model.
[0064] The execution module is used to deploy a general assembly model to drive the robot to perform human-like assembly actions.
[0065] The third objective of this invention is achieved through the following technical solution: a storage medium storing a program, which, when executed by a processor, implements the robot autonomous assembly method based on imitation learning as described in the first objective.
[0066] The fourth objective of this invention is achieved through the following technical solution: a computer device, including a processor and a memory for storing processor-executable programs, wherein when the processor executes the program stored in the memory, it implements the robot autonomous assembly method based on imitation learning as described in the first objective.
[0067] The present invention has the following advantages over the prior art:
[0068] 1. The method of the present invention constructs a universal expert strategy applicable to polygonal shaft hole assembly for data acquisition, which solves the problems of time-consuming and labor-intensive demonstration data acquisition and difficulty in skill learning caused by inconsistent demonstrations.
[0069] 2. The method of the present invention constructs a general assembly model based on shared knowledge and task-specific knowledge. This general assembly model can learn shared knowledge from demonstration data, thereby simplifying skill learning into the learning of task-specific knowledge; at the same time, it can also alleviate the problem of marginalization of information extraction, thereby realizing the learning of task-specific knowledge.
[0070] 3. This invention can overcome the problems of poor generalization performance, slow convergence and high cost of end-to-end learning methods, which are inherent in imitation learning methods. It can enable rapid learning of human skills and realize human-like autonomous robot assembly. Attached Figure Description
[0071] Figure 1 This is a flowchart illustrating the robot autonomous assembly method based on imitation learning according to the present invention.
[0072] Figure 2 This is a schematic diagram of the demonstration platform used in this invention for data acquisition. Detailed Implementation
[0073] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0074] Example 1
[0075] like Figure 1 As shown, this robot autonomous assembly method based on imitation learning includes the following steps:
[0076] S1. Analyze and summarize the expert experience in assembling polygonal shafts and holes, and construct a general expert strategy for assembling polygonal shafts and holes;
[0077] S2. Based on a general expert strategy, data is collected through demonstration to build a database; during the assembly demonstration, task-type images, assembly force / torque signals, and assembly actions are collected. Specifically, such as... Figure 2 The platform shown is used for human demonstration data collection. The specific steps are as follows:
[0078] The S21 six-degree-of-freedom industrial robot has a camera attached to its end effector to capture task-type images and collect real-time images of the expert's hand movements. At the same time, during the expert demonstration, a six-dimensional force sensor is used to measure the force and torque of the expert's hand contact during the demonstration.
[0079] S22. The images captured by the camera are transmitted to the controller via USB to calculate the position and posture of the hands during the expert demonstration;
[0080] S23. The force sensor's measured value is converted into a voltage signal by a converter, processed by the Beckhoff module and converted into a digital quantity. This digital quantity and the encoder values of each joint of the robot are transmitted to the controller via the EtherCAT protocol.
[0081] The system uses cameras and force sensors to capture the expert's assembly movements and forces / torques in real time with a sampling period of 50ms. These signals are then fed into a low-pass filter to eliminate high-frequency noise. During assembly, the expert keeps their eyes closed and adjusts their movements solely based on the perceived forces and torques.
[0082] S3. Construct a general assembly model based on shared knowledge and task-specific knowledge:
[0083] [ W v, F ω] T =p·[ W f- W f d , F τ- F τ d ] T +ζ,
[0084] Where p and ζ represent skill control parameters, which are related to the assembly task; W f d , F τ d Thresholds representing the contact states of each assembly;
[0085] Step S3 includes the following steps:
[0086] S31. Construct a Gaussian mixture regression model to learn shared knowledge from the data collected from expert demonstrations;
[0087] S32. Construct a parallel encoding attention network. This parallel encoding attention network is based on a database and outputs learned task-specific knowledge. The parallel encoding attention network takes the acquired task type image and assembly force / torque signal as input and outputs learned task-specific knowledge. Specifically, the parallel encoding attention network in step S32 includes:
[0088] The task-level encoder employs an encoder-decoder model architecture, integrating multi-layer 2D convolutional networks to extract implicit task-level information from acquired task-type images.
[0089]
[0090] in, It is the input task scene image; This indicates the extracted implicit task information; Etask (·) and D task (·) represent the task encoder function and the task decoder function, respectively; This indicates the task scenario for refactoring;
[0091] The action layer encoder adopts an encoder-decoder model architecture, integrating multiple temporal convolutional networks to extract implicit action layer information from the acquired assembly force / torque signals:
[0092]
[0093] in, It is an input force / torque signal segment; This indicates the extracted implicit action information; E action (·) and D action (·) represents the motion encoder function and the motion decoder function; This represents the reconstructed force / torque signal segment;
[0094] The attention aggregation strategy employs an attention mechanism to analyze the importance of information at different levels. It weights and aggregates information by calculating a normalized importance matrix to achieve task-specific learning.
[0095]
[0096] Where W1, W2, and W3 represent three learnable parameter matrices, and FC(·) represents a fully connected layer.
[0097] S33. Construct a multi-task meta-learning generalization framework to guide and encode attention network training in parallel, enabling rapid generalization of knowledge for specific tasks.
[0098] Step S33 includes the following steps:
[0099] S331. Multi-task learning is employed to jointly train all tasks in multi-task learning to learn the optimal parallel coding applicable to all assembly tasks. Note the network parameters; the mathematical expression for parameter updates in multi-task learning is:
[0100]
[0101] Where, θ MTL γ represents the parameters of a parallel encoded attention network trained using multi-task learning; MTL Indicates the update step size for multi-task learning; N represents the loss value for the i-th training task; task The number of tasks in the training task set; Represents task η i Supporting dataset;
[0102] S332. Meta-learning cross-task training is used to learn a set of general parallel encoding attention network parameters, which are then transferred to the new assembly task through gradient updates.
[0103] S333. Combining multi-task learning and meta-learning algorithms to achieve rapid generalization of parameters to specific task knowledge, the update formula for the combined multi-task learning-meta-learning algorithm is as follows:
[0104]
[0105] Step S332, the cross-task training process of meta-learning, includes inner loop update and outer loop update. The expression for the inner loop update is:
[0106]
[0107] in, It is based on task η i Updated parallel encoding attention network parameters; θ Meta For general parallel encoding attention network parameters trained by meta-learning; α Meta This indicates the step size for updating the inner loop; Represents task η i Supporting dataset;
[0108] The expression for updating the outer loop is:
[0109]
[0110] Where, β Meta Update the step size for the outer loop; Represents task η i The query dataset.
[0111] S34. Construct a general assembly model based on the obtained shared knowledge and the generalized specific tasks.
[0112] S4. Based on the database, a loss function-guided algorithm is used to train and optimize the general assembly model; the loss function-guided algorithm in step S4 includes a reconstruction loss function and an action error loss function.
[0113] The reconstruction error loss function is used as an optimization metric to minimize the error between the reconstructed task scenario and force / torque signal and the input task scenario and force / torque signal. Its mathematical expression is:
[0114]
[0115] in, This indicates the task scenario for refactoring; It is the input task scene image; This represents the reconstructed force / torque signal segment; It is an input force / torque signal segment;
[0116] The action error loss function, which aims to minimize the error between the predicted expert action and the actual expert action, is mathematically expressed as follows:
[0117]
[0118] in, Indicates the actions of prediction experts; W v k , F ω k ] T This indicates the actions of a real expert.
[0119] S5. Deploy the optimized general assembly model into the robot's control system to drive the robot to perform human-like assembly actions.
[0120] This embodiment addresses the problems of time-consuming and labor-intensive demonstration data acquisition and difficulties in skill learning due to demonstration inconsistencies. It constructs a general expert strategy applicable to polygonal shaft-hole assembly. Based on this strategy, human assembly skills are modeled as a general skill model based on shared knowledge and task-specific knowledge. A Gaussian mixture model is used to learn shared knowledge from demonstration data, simplifying skill learning into learning task-specific knowledge. A parallel encoding attention network is then utilized to extract and fuse implicit skill information from the task and action layers, alleviating the marginalization problem of information extraction and achieving task-specific knowledge learning. A generalization framework based on multi-task learning and meta-learning improves the rapid learning and generalization of the parallel encoding attention network in new assembly tasks.
[0121] The second objective of this invention is achieved through the following technical solution: a robot autonomous assembly system based on imitation learning, used to implement the robot autonomous assembly method based on imitation learning described in the first objective, comprising:
[0122] The strategy formulation module is used to analyze and summarize the experience of experts in polygonal shaft and hole assembly, and to construct a general expert strategy applicable to polygonal shaft and hole assembly.
[0123] The data acquisition and processing module is used to collect expert demonstration data and establish a database for training.
[0124] The model module is used to build a general assembly model based on general knowledge and task-specific knowledge, and to train and optimize the general assembly model.
[0125] The execution module is used to deploy a general assembly model to drive the robot to perform human-like assembly actions.
[0126] The third objective of this invention is achieved through the following technical solution: a storage medium storing a program, which, when executed by a processor, implements the robot autonomous assembly method based on imitation learning as described in the first objective.
[0127] The fourth objective of this invention is achieved through the following technical solution: a computer device, including a processor and a memory for storing processor-executable programs, wherein when the processor executes the program stored in the memory, it implements the robot autonomous assembly method based on imitation learning as described in the first objective.
[0128] Example 2
[0129] The robot autonomous assembly method based on imitation learning in this embodiment is the same as that in Embodiment 1, except for the following technical features:
[0130] Multi-task learning and meta-learning are combined using a mutual supervision mechanism to achieve rapid generalization of knowledge for specific tasks based on parameters. The expression for this is:
[0131]
[0132] Where, θ MTL θ represents the parameters of a parallel encoded attention network trained using multi-task learning; Meta The basic HIL parameters for meta-learning training; γ MTL Indicates the update step size for multi-task learning; N represents the loss value for the i-th training task; task The number of tasks in the training task set; Represents task η i Supporting datasets; EE * The expected number of training iterations during parameter backtracking is represented by EE; EE represents the current number of training iterations.
[0133] This embodiment combines multi-task learning and meta-learning through a mutual supervision mechanism, which avoids the problem of negative knowledge transfer in the learning process and has strong generalization ability.
[0134] The above-described specific embodiments are preferred embodiments of the present invention and are not intended to limit the present invention. Any other changes or equivalent substitutions made without departing from the technical solution of the present invention are included within the protection scope of the present invention.
Claims
1. A robot autonomous assembly method based on imitation learning, characterized in that, Includes the following steps: S1. Analyze and summarize the expert experience in assembling polygonal shafts and holes, and construct a general expert strategy for assembling polygonal shafts and holes; S2. Based on a general expert strategy, data is collected through demonstration to build a database; S3. Construct a general assembly model based on shared knowledge and task-specific knowledge: ， wherein, , denotes a skill control parameter related to the assembly task; , denotes a threshold value for the respective assembly contact state; Step S3 includes the following steps: S31. Construct a Gaussian mixture regression model to learn shared knowledge from the data collected from expert demonstrations; S32. Construct a parallel encoding attention network. This parallel encoding attention network is based on a database and outputs the learned task-specific knowledge. S33. Construct a multi-task meta-learning generalization framework to guide the training of parallel encoding attention networks and achieve rapid generalization of knowledge for specific tasks. Step S33 includes the following steps: S331. Multi-task learning is employed to jointly train all tasks in multi-task learning to learn the optimal parallel encoding attention network parameters applicable to all assembly tasks. The mathematical expression for parameter update in multi-task learning is: wherein, denotes a parallel encoding attention network parameter trained with multi-task learning; denotes an update step size of multi-task learning; is a loss value of the i-th training task; is a number of tasks of the training task set; denotes a support data set of the task . S332. Meta-learning cross-task training is used to learn a set of general parallel encoding attention network parameters, which are then transferred to the new assembly task through gradient updates. S333. Combining multi-task learning and meta-learning algorithms to achieve rapid generalization of parameters to specific task knowledge, the update formula for the combined multi-task learning-meta-learning algorithm is as follows: ； or, Multi-task learning and meta-learning are combined using a mutual supervision mechanism to achieve rapid generalization of knowledge for specific tasks based on parameters. The expression for this is: in, This represents the parameters of a parallel encoded attention network trained using multi-task learning; These are the basic HIL parameters for meta-learning training; Indicates the update step size for multi-task learning; Let be the loss value for the i-th training task; The number of tasks in the training task set; Indicates task Supporting dataset; This represents the number of training iterations required for the expected parameter backhaul. Indicates the current number of training iterations; S34. Construct a general assembly model based on the obtained shared knowledge and the generalized specific task knowledge; S4. Based on the database, a loss function is used to guide the algorithm training and optimization of the general assembly model; S5. Deploy the optimized general assembly model into the robot's control system to drive the robot to perform human-like assembly actions.
2. The robot autonomous assembly method based on imitation learning according to claim 1, characterized in that, Step S2 includes the following steps: The S21 six-degree-of-freedom industrial robot has a camera attached to its end effector to capture task-type images and collect real-time images of the expert's hand movements. At the same time, during the expert's demonstration, a six-dimensional force sensor is used to measure the force and torque of the expert's hand contact during the demonstration. S22. The images captured by the camera are transmitted to the controller via USB to calculate the position and posture of the hands during the expert demonstration; S23. The force sensor's measured value is converted into a voltage signal by a converter, processed by the Beckhoff module and converted into a digital quantity. This digital quantity and the encoder values of each joint of the robot are transmitted to the controller via the EtherCAT protocol.
3. The robot autonomous assembly method based on imitation learning according to claim 1, characterized in that, The parallel encoding attention network mentioned in step S32 includes: The task-level encoder employs an encoder-decoder model architecture, integrating multi-layer 2D convolutional networks to extract implicit task-level information from acquired task-type images. in, It is the input task scene image; This indicates the extracted implicit task information; and These represent the task encoder function and the task decoder function, respectively. This indicates the task scenario for refactoring; The action layer encoder adopts an encoder-decoder model architecture, integrating multiple temporal convolutional networks to extract implicit action layer information from the acquired assembly force / torque signals: in, It is an input force / torque signal segment; This indicates the extracted implicit action information; and This represents the motion encoder function and the motion decoder function; This represents the reconstructed force / torque signal segment; The attention aggregation strategy employs an attention mechanism to analyze the importance of information at different levels. It weights and aggregates information by calculating a normalized importance matrix to achieve task-specific learning. Where W1, W2, and W3 represent three learnable parameter matrices, This indicates a fully connected layer.
4. The robot autonomous assembly method based on imitation learning according to claim 1, characterized in that, Step S332, the cross-task training process of meta-learning, includes inner loop update and outer loop update. The expression for updating the inner loop is: in, It is based on the task Updated parameters for the parallel encoding attention network; These are the parameters for a general parallel encoding attention network trained by meta-learning; This indicates the step size for updating the inner loop; Indicates task Supporting dataset; The expression for updating the outer loop is: in, Update the step size for the outer loop; Indicates task The query dataset.
5. The robot autonomous assembly method based on imitation learning according to claim 1, characterized in that, The loss function guiding the algorithm in step S4 includes the reconstruction error loss function and the action error loss function. The reconstruction error loss function is used as an optimization metric to minimize the error between the reconstructed task scenario and force / torque signal and the input task scenario and force / torque signal. Its mathematical expression is: in, This indicates the task scenario for refactoring; It is the input task scene image; This represents the reconstructed force / torque signal segment; It is an input force / torque signal segment; The action error loss function, with the optimization objective of minimizing the error between the predicted expert action and the actual expert action, is expressed mathematically as follows: in, This indicates the actions of forecasting experts; This indicates the actions of a real expert.
6. A robot autonomous assembly system based on imitation learning, characterized in that, The method for implementing the robot autonomous assembly method based on imitation learning as described in any one of claims 1 to 5 includes: The strategy formulation module is used to analyze and summarize the experience of experts in polygonal shaft and hole assembly, and to construct a general expert strategy applicable to polygonal shaft and hole assembly. The data acquisition and processing module is used to collect expert demonstration data and establish a database for training. The model module is used to build a general assembly model based on shared knowledge and task-specific knowledge, and to train and optimize the general assembly model. The execution module is used to deploy a general assembly model to drive the robot to perform human-like assembly actions.
7. A storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the robot autonomous assembly method based on imitation learning as described in any one of claims 1 to 5.
8. A computer device, comprising a processor and a memory for storing a processor-executable program, characterized in that, When the processor executes the program stored in the memory, it implements the robot autonomous assembly method based on imitation learning as described in any one of claims 1 to 5.