Robot autonomous assembly method and system based on imitation learning, storage medium and computer equipment

Abstract

The invention discloses a robot autonomous assembly method and system based on imitation learning, a storage medium and computer equipment, and the method comprises the steps: constructing a universal expert strategy suitable for polygonal shaft hole assembly through analyzing and summarizing the polygonal shaft hole assembly experience of an expert; according to the strategy, task type images and human assembly demonstration data including assembly force / torque and assembly actions are collected; modeling a human assembly skill into a universal assembly skill model based on shared knowledge and specific knowledge; using a loss function guidance algorithm to train and optimize the general assembly model; and deploying the optimized general assembly model into a control system of the robot so as to drive the robot to execute humanization-like assembly. According to the method, the problems of poor generalization performance, slow convergence of an end-to-end learning mode and high cost of an imitation learning method can be solved, rapid learning of human skills can be realized, and humanoid robot autonomous assembly is realized.

Description

Technical Field

[0001] The present invention relates to robotics technology, and in particular to a robot autonomous assembly method, system, storage medium and computer equipment based on imitation learning. Background Art

[0002] The essence of robot assembly skill learning is to learn the expert's strategy of adjusting assembly actions based on perceived force / torque. Currently, policy learning methods are mainly divided into two categories. The first is policy learning methods based on reinforcement learning frameworks. These methods can achieve autonomous learning by interacting with various assembly environments without prior knowledge. The second is policy learning methods based on supervised learning frameworks. These methods summarize human assembly strategies from human demonstrations. Compared to reinforcement learning algorithms, imitation learning algorithms do not require interaction with the environment, avoiding the exploration and trial-and-error process, and can therefore learn and imitate human strategies more quickly. However, imitation learning models often lack the ability to generalize to unseen environments or situations. Performance degrades in new tasks.

[0003] Research has shown that meta-learning can improve the model's ability to generalize quickly on data-scarce tasks through cross-task training. Among them, Model-Agnostic Meta-Learning (MAML) is a state-of-the-art representative meta-learning algorithm that has been widely used in various fields. The key idea of ​​MAML is to learn the best initial model parameters through cross-task training to help the model quickly adapt to new tasks with only a small amount of gradient updates. MAML's task independence allows it to be flexibly applied to any model learned by gradient descent. However, MAML has the problem that training requires a large number of similar meta-tasks and the training convergence is slow. In addition, for tasks with obvious differences in the distribution of training data, MAML finds it difficult to learn effective shallow features, which is related to the performance of meta-learning.

[0004] To address these issues, meta-transfer learning algorithms are being researched. These algorithms leverage transfer learning to transfer models pre-trained on multi-task datasets to meta-learning algorithms, improving convergence and generalization. Research is also underway to combine MAML with transfer learning algorithms to improve generalization. While these methods can accelerate algorithm convergence, transfer learning suffers from the problem of negative transfer. Achieving rapid skill learning and generalization with limited data remains challenging. Summary of the Invention

[0005] The present invention aims to overcome the shortcomings of the above-mentioned prior art and provides 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 difficulty in skill learning caused by inconsistent demonstrations.

[0006] The second object of the present invention is to provide a robot autonomous assembly system based on imitation learning.

[0007] A third object of the present invention is to provide a storage medium.

[0008] A fourth object of the present invention is to provide a computer device.

[0009] The first object of the present invention is achieved by the following technical solution: This robot autonomous assembly method based on imitation learning comprises the following steps:

[0010] S1. Analyze and summarize the expert polygonal shaft hole assembly experience and construct a universal expert strategy for polygonal shaft hole assembly;

[0011] S2, based on universal expert strategies, data collection is performed through demonstration to build a database;

[0012] S3. Build 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] Among them, p,ζ represents the skill control parameters, which are related to the assembly task; W f d , F τ d Thresholds representing contact status of each assembly;

[0015] S4. Based on the database, a loss function is used to guide the algorithm to train and optimize 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 movements.

[0017] The step S2 comprises the following steps:

[0018] S21. A camera is connected to the end of the six-degree-of-freedom industrial robot to capture task type conversion images and real-time capture 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.

[0019] S22, the images captured by the camera are transmitted to the controller via USB to calculate the position and posture of the expert's hand during the demonstration;

[0020] S23. The measured value of the force sensor 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 value 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 data collected from expert demonstrations.

[0023] S32. Build a parallel coding attention network, which outputs the learned task-specific knowledge based on the database;

[0024] S33. Build a multi-task meta-learning generalization framework to guide and parallelize the training of attention networks to achieve rapid generalization of task-specific knowledge.

[0025] S34. Construct a general assembly model based on the obtained shared knowledge and generalized specific tasks.

[0026] The parallel encoding attention network in step S32 includes:

[0027] The task-level encoder uses an encoder-decoder model architecture and integrates a multi-layer 2D convolutional network to extract implicit task-level information from the collected task-type images:

[0028]

[0029] in, is the input task scene picture; represents the extracted implicit task information; E task (·) and D task (·) represents the task encoder function and the task decoder function respectively; Represents the reconstructed task scenario;

[0030] The action layer encoder uses an encoder-decoder model architecture and integrates a multi-layer temporal convolutional network to extract implicit action layer information from the collected assembly force / torque signals:

[0031]

[0032] in, is the input force / torque signal segment; represents the extracted implicit action information; E action(·) and D action (·) represents the action encoder function and the action decoder function; represents the reconstructed force / torque signal segment;

[0033] The attention aggregation strategy uses the attention mechanism to analyze the importance of information at different levels and weights the information by calculating the normalized importance matrix to achieve specific task learning:

[0034]

[0035] Among them, W1, W2, and W3 represent three learnable parameter matrices, and FC(·) represents the fully connected layer.

[0036] Step S33 includes the following steps:

[0037] S331. Multi-task learning is used 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 of multi-task learning is:

[0038]

[0039] Among them, θ MTL represents the parameters of the parallel encoding attention network trained using multi-task learning; γ MTL represents the update step size of multi-task learning; is the loss value of the i-th training task; N task is the number of tasks in the training task set; Represents task η i Supporting datasets;

[0040] S332. Meta-learning cross-task training is used to learn a common set of parallel encoding attention network parameters, which are transferred to the new assembly task through gradient updates.

[0041] S333. Combine multi-task learning and meta-learning algorithms to achieve rapid generalization of task-specific knowledge for parameters. The update formula for the combination of multi-task learning and meta-learning algorithms is:

[0042]

[0043] or,

[0044] Multi-task learning and meta-learning are combined through mutual supervision to achieve rapid generalization of task-specific knowledge for parameters. The expression is:

[0045]

[0046] Among them, among them, θMTL represents the parameters of the parallel encoding attention network trained using multi-task learning; θ Meta is the basic HIL parameter for meta-learning training; γ MTL represents the update step size of multi-task learning; is the loss value of the i-th training task; N task is the number of tasks in the training task set; Represents task η i Supporting datasets for EE * Indicates the expected number of training times for parameter migration; EE indicates the current number of training times.

[0047] The meta-learning cross-task training process in step S332 includes inner loop update and outer loop update. The expression of the inner loop update is:

[0048]

[0049] in, According to the task η i Updated parallel encoding attention network parameters; θ Meta General parallel encoding attention network parameters for meta-learning training; α Meta Indicates the inner loop update step size; Represents task η i Supporting datasets;

[0050] The expression for the outer loop update is:

[0051]

[0052] Among them, β Meta Update the step size for the outer loop; Represents task η i The query dataset.

[0053] The loss function guidance algorithm in step S4 includes the reconstruction loss function and the action error loss function.

[0054] The reconstruction error loss function is used to minimize the error between the reconstructed task scene and force / torque signal and the input task scene and force / torque signal as the optimization indicator. Its mathematical expression is:

[0055]

[0056] in, Represents the reconstructed task scenario; is the input task scene picture; represents the reconstructed force / torque signal segment; is the input force / torque signal segment;

[0057] The action error loss function is used to minimize the error between the predicted expert action and the real expert action as the optimization goal. Its mathematical expression is:

[0058]

[0059] in, represents the predicted expert action; [ W v k , F ω k ] T Indicates real expert actions.

[0060] The second object of the present 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 object, comprising:

[0061] Strategy formulation module, used to analyze and summarize the expert polygonal shaft hole assembly experience and construct a universal expert strategy applicable to polygonal shaft hole assembly;

[0062] Data acquisition and processing module, used to collect expert demonstration data and build a database for training;

[0063] The model module is used to build a general assembly model based on general knowledge and specific task 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 object of the present invention is achieved through the following technical solution: a storage medium stores a program, and when the program is executed by a processor, the robot autonomous assembly method based on imitation learning described in the first object is achieved.

[0066] The fourth purpose of the present invention is achieved through the following technical solution: a computer device includes a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the robot autonomous assembly method based on imitation learning described in the first purpose is achieved.

[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 suitable for polygonal shaft hole assembly for data collection, which solves the problems of time-consuming and labor-intensive demonstration data collection and difficulty in skill learning caused by inconsistent demonstrations.

[0069] 2. The method of the present invention constructs a universal assembly model based on shared knowledge and specific task knowledge. This universal assembly model can learn shared knowledge from demonstration data, thereby simplifying skill learning into learning specific task knowledge; at the same time, it can also alleviate the marginalization problem of information extraction, thereby realizing specific task knowledge learning.

[0070] 3. The present invention can overcome the problems of poor generalization performance of imitation learning methods, slow convergence of end-to-end learning methods, and high costs, and can achieve rapid learning of human skills and realize human-like robot autonomous assembly. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a flow chart of the robot autonomous assembly method based on imitation learning of the present invention.

[0072] Figure 2 It is a schematic diagram of a demonstration platform used when the present invention performs data collection. DETAILED DESCRIPTION

[0073] The present invention will be further described below with reference to the accompanying drawings and examples.

[0074] Example 1

[0075] like Figure 1 As shown, the robot autonomous assembly method based on imitation learning includes the following steps:

[0076] S1. Analyze and summarize the expert polygonal shaft hole assembly experience and construct a universal expert strategy for polygonal shaft hole assembly;

[0077] S2. Based on the universal expert strategy, data is collected through demonstration to build a database; task type images, assembly force / torque signals, and assembly actions are collected during the demonstration assembly process. Specifically, Figure 2 The platform shown in the figure is used to collect human demonstration data. The specific steps are as follows:

[0078] S21. A camera is connected to the end of the six-degree-of-freedom industrial robot to capture task type conversion images and real-time capture 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.

[0079] S22, the images captured by the camera are transmitted to the controller via USB to calculate the position and posture of the expert's hand during the demonstration;

[0080] S23. The measured value of the force sensor 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 value of each joint of the robot are transmitted to the controller via the EtherCAT protocol.

[0081] The camera and force sensor capture the expert's assembly movements and assembly forces / torques in real time with a 50ms sampling period. These signals are fed into a low-pass filter to eliminate high-frequency noise. During the assembly process, the human operator closes their eyes and adjusts their movements based solely on the perceived forces and torques.

[0082] S3. Build 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] Among them, p,ζ represents the skill control parameters, which are related to the assembly task; W f d , F τ d Thresholds representing contact status of each assembly;

[0085] Step S3 includes the following steps:

[0086] S31. Construct a Gaussian mixture regression model to learn shared knowledge from data collected from expert demonstrations.

[0087] S32. Construct a parallel coding attention network. This parallel coding attention network outputs the learned specific task knowledge based on the database. The parallel coding attention network takes the collected task type image and assembly force / torque signal as input and outputs the learned specific task knowledge. Specifically, the parallel coding attention network in step S32 includes:

[0088] The task-level encoder uses an encoder-decoder model architecture and integrates a multi-layer 2D convolutional network to extract implicit task-level information from the collected task-type images:

[0089]

[0090] in, is the input task scene picture; represents the extracted implicit task information; Etask (·) and D task (·) represents the task encoder function and the task decoder function respectively; Represents the reconstructed task scenario;

[0091] The action layer encoder uses an encoder-decoder model architecture and integrates a multi-layer temporal convolutional network to extract implicit action layer information from the collected assembly force / torque signals:

[0092]

[0093] in, is the input force / torque signal segment; represents the extracted implicit action information; E action (·) and D action (·) represents the action encoder function and the action decoder function; represents the reconstructed force / torque signal segment;

[0094] The attention aggregation strategy uses the attention mechanism to analyze the importance of information at different levels and weights the information by calculating the normalized importance matrix to achieve specific task learning:

[0095]

[0096] Among them, W1, W2, and W3 represent three learnable parameter matrices, and FC(·) represents the fully connected layer.

[0097] S33. Build a multi-task meta-learning generalization framework to guide and parallelize the training of attention networks to achieve rapid generalization of task-specific knowledge.

[0098] Step S33 includes the following steps:

[0099] S331. Multi-task learning is used 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 of multi-task learning is:

[0100]

[0101] Among them, θ MTL represents the parameters of the parallel encoding attention network trained using multi-task learning; γ MTL represents the update step size of multi-task learning; is the loss value of the i-th training task; N task is the number of tasks in the training task set; Represents task η i Supporting datasets;

[0102] S332. Meta-learning cross-task training is used to learn a common set of parallel encoding attention network parameters, which are transferred to the new assembly task through gradient updates.

[0103] S333. Combine multi-task learning and meta-learning algorithms to achieve rapid generalization of task-specific knowledge for parameters. The update formula for the combination of multi-task learning and meta-learning algorithms is:

[0104]

[0105] The meta-learning cross-task training process in step S332 includes inner loop update and outer loop update. The expression of the inner loop update is:

[0106]

[0107] in, According to the task η i Updated parallel encoding attention network parameters; θ Meta General parallel encoding attention network parameters for meta-learning training; α Meta Indicates the inner loop update step size; Represents task η i Supporting datasets;

[0108] The expression for the outer loop update is:

[0109]

[0110] Among them, 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 generalized specific tasks.

[0112] S4, based on the database, adopt the loss function guidance algorithm to train and optimize the general assembly model; the loss function guidance algorithm in step S4 includes the reconstruction loss function and the action error loss function,

[0113] The reconstruction error loss function is used to minimize the error between the reconstructed task scene and force / torque signal and the input task scene and force / torque signal as the optimization indicator. Its mathematical expression is:

[0114]

[0115] in, Represents the reconstructed task scenario; is the input task scene picture; represents the reconstructed force / torque signal segment; is the input force / torque signal segment;

[0116] The action error loss function is used to minimize the error between the predicted expert action and the real expert action as the optimization goal. Its mathematical expression is:

[0117]

[0118] in, represents the predicted expert action; [ W v k , F ω k ] T Indicates real expert actions.

[0119] S5. Deploy the optimized general assembly model into the robot's control system to drive the robot to perform human-like assembly movements.

[0120] The method of this embodiment addresses the problems of time-consuming and labor-intensive demonstration data collection and difficulty in skill learning caused by demonstration inconsistency, and constructs a universal expert strategy suitable for polygonal shaft hole assembly. Based on this strategy, human assembly skills are modeled as a universal skill model based on shared knowledge and task-specific knowledge, and a Gaussian mixture model is used to learn shared knowledge from demonstration data, thereby simplifying skill learning to the learning of task-specific knowledge. The parallel coding attention network is then used to extract and fuse the implicit skill information of the task layer and the action layer, alleviating the marginalization problem of information extraction, thereby realizing task-specific knowledge learning. The generalization framework based on multi-task learning and meta-learning improves the rapid learning and generalization of the parallel coding attention network in new assembly tasks.

[0121] The second object of the present 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 object, comprising:

[0122] Strategy formulation module, used to analyze and summarize the expert polygonal shaft hole assembly experience and construct a universal expert strategy applicable to polygonal shaft hole assembly;

[0123] Data acquisition and processing module, used to collect expert demonstration data and build a database for training;

[0124] The model module is used to build a general assembly model based on general knowledge and specific task 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 object of the present invention is achieved through the following technical solution: a storage medium stores a program, and when the program is executed by a processor, the robot autonomous assembly method based on imitation learning described in the first object is achieved.

[0127] The fourth purpose of the present invention is achieved through the following technical solution: a computer device includes a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the robot autonomous assembly method based on imitation learning described in the first purpose is achieved.

[0128] Example 2

[0129] The robot autonomous assembly method based on imitation learning in this embodiment is the same as that in Example 1 except for the following technical features:

[0130] Multi-task learning and meta-learning are combined through mutual supervision to achieve rapid generalization of task-specific knowledge for parameters. The expression is:

[0131]

[0132] Among them, θ MTL represents the parameters of the parallel encoding attention network trained using multi-task learning; θ Meta is the basic HIL parameter for meta-learning training; γ MTL represents the update step size of multi-task learning; is the loss value of the i-th training task; N task is the number of tasks in the training task set; Represents task η i Supporting datasets for EE * Indicates the expected number of training times for parameter migration; EE indicates the current number of training times.

[0133] In this embodiment, multi-task learning and meta-learning are combined using a mutual supervision mechanism, which can avoid the problem of negative knowledge transfer in the learning process and has strong generalization ability.

[0134] The above specific implementation manner is a preferred embodiment of the present invention and does not limit the present invention. Any other changes or other equivalent replacement methods that do not deviate from the technical solution of the present invention are included in the protection scope of the present invention.

Claims

1. A robot autonomous assembly method based on imitation learning, characterized in that: The following steps are involved: S1. Analyze and summarize the expert polygonal shaft hole assembly experience and construct a universal expert strategy for polygonal shaft hole assembly; S2, based on universal expert strategies, data collection is performed through demonstration to build a database; S3. Build a general assembly model based on shared knowledge and task-specific knowledge: [ W v, F oh] T =p·[ W f- W f d , F t- F t d ] T +ζ, Among them, p,ζ represents the skill control parameters, which are related to the assembly task; W f d , F τ d Thresholds representing contact status of each assembly; S4. Based on the database, a loss function is used to guide the algorithm to train and optimize 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 movements.

2. The robot autonomous assembly method based on imitation learning according to claim 1, characterized in that: The step S2 comprises the following steps: S21. A camera is connected to the end of the six-degree-of-freedom industrial robot to capture task type conversion images and real-time capture 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 expert's hand during the demonstration; S23. The measured value of the force sensor 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 value 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: Step S3 includes the following steps: S31. Construct a Gaussian mixture regression model to learn shared knowledge from data collected from expert demonstrations. S32. Build a parallel coding attention network, which outputs the learned task-specific knowledge based on the database; S33. Build a multi-task meta-learning generalization framework to guide and parallelize the training of attention networks to achieve rapid generalization of task-specific knowledge. S34. Construct a general assembly model based on the obtained shared knowledge and generalized specific tasks.

4. The robot autonomous assembly method based on imitation learning according to claim 3 is characterized in that: The parallel encoding attention network in step S32 includes: The task-level encoder uses an encoder-decoder model architecture and integrates a multi-layer 2D convolutional network to extract implicit task-level information from the collected task-type images: in, is the input task scene picture; represents the extracted implicit task information; E task (·) and D task (·) represents the task encoder function and the task decoder function respectively; Represents the reconstructed task scenario; The action layer encoder uses an encoder-decoder model architecture and integrates a multi-layer temporal convolutional network to extract implicit action layer information from the collected assembly force / torque signals: in, is the input force / torque signal segment; represents the extracted implicit action information; E action (·) and D action (·) represents the action encoder function and the action decoder function; represents the reconstructed force / torque signal segment; The attention aggregation strategy uses the attention mechanism to analyze the importance of information at different levels and weights the information by calculating the normalized importance matrix to achieve specific task learning: Among them, W1, W2, and W3 represent three learnable parameter matrices, and FC(·) represents the fully connected layer.

5. The robot autonomous assembly method based on imitation learning according to claim 3 is characterized in that: Step S33 includes the following steps: S331. Multi-task learning is used 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 of multi-task learning is: Among them, θ MTL represents the parameters of the parallel encoding attention network trained using multi-task learning; γ MTL represents the update step size of multi-task learning; is the loss value of the i-th training task; N task is the number of tasks in the training task set; Represents task η i Supporting datasets; S332. Meta-learning cross-task training is used to learn a common set of parallel encoding attention network parameters, which are transferred to the new assembly task through gradient updates. S333. Combine multi-task learning and meta-learning algorithms to achieve rapid generalization of task-specific knowledge for parameters. The update formula for the combination of multi-task learning and meta-learning algorithms is: or, Multi-task learning and meta-learning are combined through mutual supervision to achieve rapid generalization of task-specific knowledge for parameters. The expression is: Among them, θ MTL represents the parameters of the parallel encoding attention network trained using multi-task learning; θ Meta is the basic HIL parameter for meta-learning training; γ MTL represents the update step size of multi-task learning; is the loss value of the i-th training task; N task is the number of tasks in the training task set; Represents task η i Supporting datasets for EE * Indicates the expected number of training times for parameter migration; EE indicates the current number of training times.

6. The robot autonomous assembly method based on imitation learning according to claim 5, characterized in that: The meta-learning cross-task training process in step S332 includes inner loop update and outer loop update. The expression for the inner loop update is: in, According to the task η i Updated parallel encoding attention network parameters; θ Meta General parallel encoding attention network parameters for meta-learning training; α Meta Indicates the inner loop update step size; Represents task η i Supporting datasets; The expression for the outer loop update is: Among them, β Meta Update the step size for the outer loop; Represents task η i The query dataset.

7. The robot autonomous assembly method based on imitation learning according to claim 1, characterized in that: The loss function guidance algorithm in step S4 includes the reconstruction loss function and the action error loss function. The reconstruction error loss function is used to minimize the error between the reconstructed task scene and force / torque signal and the input task scene and force / torque signal as the optimization indicator. Its mathematical expression is: in, Represents the reconstructed task scenario; is the input task scene picture; represents the reconstructed force / torque signal segment; is the input force / torque signal segment; The action error loss function is used to minimize the error between the predicted expert action and the real expert action as the optimization goal. Its mathematical expression is: in, represents the predicted expert action; Indicates real expert actions.

8. The robot autonomous assembly system based on imitation learning is characterized by: The method for implementing the robot autonomous assembly method based on imitation learning according to any one of claims 1 to 7 comprises: Strategy formulation module, used to analyze and summarize the expert polygonal shaft hole assembly experience and construct a universal expert strategy applicable to polygonal shaft hole assembly; Data acquisition and processing module, used to collect expert demonstration data and build a database for training; The model module is used to build a general assembly model based on general knowledge and specific task 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.

9. A storage medium storing a program, characterized in that: When the program is executed by a processor, the robot autonomous assembly method based on imitation learning according to any one of claims 1 to 7 is implemented.

10. A computer device comprising a processor and a memory for storing a program executable by the processor, characterized in that When the processor executes the program stored in the memory, the robot autonomous assembly method based on imitation learning according to any one of claims 1 to 7 is implemented.

Images