Robot motion strategy interpretability identification method based on sparse auto-encoder

Abstract

The invention discloses a robot motion strategy interpretability identification method based on a sparse auto-encoder, and the method comprises the steps: the training of a motion strategy of a mobile robot and the training of an SAE layer of the sparse auto-encoder: employing the motion of a two-wheel differential robot as a typical robot navigation task; a simulation learning model trained on the task is used as a typical robot strategy model, and then an SAE layer is trained to analyze the strategy model as a typical normal form of strategy model analysis; extracting and explaining sparse neuron features in the strategy model: screening out effective features of a neural network by using a filtering mode, and explaining the effective features; and verifying the interpretability by activating the patch intervention strategy model. According to the embodiment of the invention, the SAE layer is adopted to decouple the superimposed effect in the neural network, so that the characteristics of neurons are relatively single, and the relevance between a neuron mode and a motion strategy can be analyzed conveniently; by activating patch intervention, the model interpretation effect is verified, and the transparency of the neural network strategy is improved.

Description

Technical Field

[0001] This invention relates to the field of robot strategy interpretability technology, and more particularly to a robot motion strategy interpretability identification method based on sparse autoencoders. Background Technology

[0002] Currently, robotic systems widely employ intelligent strategies centered on neural networks, demonstrating superior performance in complex tasks such as autonomous driving, drone navigation, and robotic arm operations. However, because neural networks rely on gradient descent training, their internal representations are highly complex, exhibiting "black box" characteristics that make it difficult for humans to directly understand their decision-making logic. This characteristic limits the effective assessment of strategy reliability in risk-sensitive scenarios (such as spatial operations), making it difficult to ensure the safety and stability of the system at critical moments. In human-machine collaborative scenarios (such as autonomous driving), the lack of interpretability makes it difficult for humans to accurately judge the system state or the timing of intervention, reducing collaborative efficiency. Furthermore, the high-dimensional characteristics of neural networks and the complex interactions between neurons further increase the difficulty of analysis, and existing interpretation methods are insufficient to meet the needs of practical applications.

[0003] Therefore, there is an urgent need for an efficient and interpretable method to improve the transparency, reliability, and human-robot collaboration capabilities of robot strategies by analyzing the intrinsic mechanisms of neural networks, thus laying the foundation for further applications of robots in the real world.

[0004] Interpretable methods in deep learning provide a general approach to these types of problems. Many interpretable methods rely on the analysis of the activation characteristics and attention matrix of neural networks, but the high dimensionality of neural network layers and the superposition of neurons make the analysis process very difficult.

[0005] Therefore, it is necessary to improve one or more of the problems existing in the above-mentioned related technical solutions.

[0006] It should be noted that this section is intended to provide background or context for the technical solutions of this disclosure as set forth in the claims. The description herein does not constitute an admission that it is prior art simply because it is included in this section. Summary of the Invention

[0007] The purpose of this invention is to provide a robot motion strategy interpretability identification method based on sparse autoencoders, thereby overcoming at least to some extent one or more problems caused by the limitations and defects of related technologies.

[0008] This invention first provides a method for identifying the interpretability of robot motion strategies based on sparse autoencoders, including: S1, Training of mobile robot motion strategy and training of sparse autoencoder (SAE) layer: Two-wheel differential robot motion is used as a typical robot navigation task, and the imitation learning model trained on this task is used as a typical robot policy model. Then, the SAE layer is trained to analyze the policy model, which serves as a typical paradigm for policy model analysis. S2, Feature extraction and interpretation of sparse neurons in the policy model: using filtering to select effective features of the neural network and interpreting them; S3 verifies interpretability by activating the patch intervention strategy model.

[0009] In this invention, in S1, the robot strategy model is an MLP network model with 9 input dimensions, 64 hidden layers, and 2 output dimensions.

[0010] In this invention, the MLP network model includes a sparse encoder network and a sparse decoder network, which are represented as follows:

[0011]

[0012] in, The values ​​of the hidden layer of the original MLP network are used as input to the SAE module. Yes The estimated value of the reconstruction is used as the output of the SAE module. , These are the linear mapping parameters to be learned by the sparse encoder, and the JumpReLU function is the activation function used in SAE. This represents the sparse characterization of SAE. It is the parameter matrix of the sparse decoder in SAE (Sparse Autoencoder). Column vectors.

[0013] In this invention, in S2, filtering is performed by multiplying the sparse representation with the dictionary.

[0014] In this invention, the following formula is satisfied:

[0015] Then it is considered that the dictionary Related to this movement pattern, This is the dictionary activation threshold; there may be multiple related dictionaries.

[0016] In this invention, dictionary features are used. Sparse features of SAE are constructed as coefficients and introduced into the original neural network to observe whether the output motion pattern of the neural network meets expectations.

[0017] The present invention further provides a robot motion strategy interpretability identification device based on a sparse autoencoder, comprising: The training module is used to train the motion strategy of the mobile robot and the Sparse Autoencoder (SAE) layer: Two-wheel differential robot motion is used as a typical robot navigation task, and the imitation learning model trained on this task is used as a typical robot policy model. The SAE layer is then trained to analyze the policy model, which serves as a typical paradigm for policy model analysis. Feature extraction and interpretation, used for feature extraction and interpretation of sparse neurons in policy models: using filtering to select effective features of the neural network and interpreting them; The verification module is used to verify interpretability by activating the patch intervention strategy model.

[0018] The technical solution provided by this invention may include the following beneficial effects: This invention presents a robot motion strategy interpretability identification method based on sparse autoencoders. It employs SAE layers to decouple the superposition effect in the neural network, making the features of neurons more ambiguous and facilitating the analysis of the correlation between neuron patterns and motion strategies. By activating patch intervention, the effectiveness of model interpretation is verified, and the transparency of neural network strategies is improved. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0020] Figure 1 A flowchart illustrating the robot motion strategy interpretability identification method based on sparse autoencoders in an exemplary embodiment of this disclosure is shown. Figure 2 This illustrates a two-wheeled differential robot simulation environment rendered by pygame in an exemplary embodiment of this disclosure; Figure 3 Examples of robot neural network strategies and SAE layers in exemplary embodiments of this disclosure are shown; Figure 4 This illustration shows a visualized robot motion trajectory and SAE activation map in an exemplary embodiment of this disclosure; Figure 5 This embodiment of the present disclosure visualizes another robot motion trajectory and SAE activation map; Figure 6 The diagram illustrates the motion trajectory of a robot under the action of an activated patch in an exemplary embodiment of this disclosure. Detailed Implementation

[0021] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0022] Furthermore, the accompanying drawings are merely illustrative diagrams of embodiments of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0023] This example implementation first provides a robot motion strategy interpretability identification method based on a sparse autoencoder. Please refer to [link / reference]. Figure 2 This method may include: S1-S3. Specifically: S1, Training of mobile robot motion strategy and training of sparse autoencoder (SAE) layer: Two-wheel differential robot motion is used as a typical robot navigation task, and the imitation learning model trained on this task is used as a typical robot policy model. Then, the SAE layer is trained to analyze the policy model, which serves as a typical paradigm for policy model analysis. S2, Feature extraction and interpretation of sparse neurons in the policy model: using filtering to select effective features of the neural network and interpreting them; S3 verifies interpretability by activating the patch intervention strategy model.

[0024] In this embodiment, the superposition effect in the neural network is decoupled using the SAE layer, making the features of neurons more ambiguous and facilitating the analysis of the correlation between neuron patterns and motion strategies. By intervening with activation patches, the effectiveness of the model interpretation is verified, and the transparency of the neural network strategy is improved.

[0025] The specific process of each step in the above embodiments will be described below.

[0026] S1, training of the mobile robot's motion strategy and SAE layer training.

[0027] This invention uses two-wheeled differential robot motion as a typical robot navigation task, and takes the imitation learning model trained on this task as a typical robot policy model. Then, the SAE layer is trained to analyze the policy model, which serves as a typical paradigm for policy model analysis.

[0028] Specifically, this invention develops a simulation environment based on the dynamics of a differential robot, where the dynamic model is as follows:

[0029] in Position vector , For angle, For quality, For rotational inertia, The driving force for the robot's current orientation. This represents the torque acting on the robot. In this simulation environment, its state is its position. ,represent Quaternions of angles ,speed and angular velocity It has a total of 9 dimensions, and its action space is the driving force. and torque Two dimensions. The robot's task objective is to move its position from its initial location to a target area near the origin of the coordinate system through actions output by the policy. In order to facilitate demonstration, a simulation environment rendering was also written using pygame, such as... Figure 2 As shown, the pink square area is the target area that the robot needs to reach, and the red trajectory lines from light to dark represent the robot's movement trajectory in the current round.

[0030] Furthermore, to construct the robot's neural network strategy, it is necessary to collect data from its successful operation as a training set for imitation learning. Therefore, a PD control law for the robot was designed to automatically collect data. The PD control law was constructed along two dimensions: the angle of the robot's head towards the origin and the distance, to achieve decoupled tracking control. One hundred trajectory data points were collected to construct the dataset. Finally, an MLP network with 9 input dimensions, 64 hidden layers, and 2 output dimensions was built as the robot's strategy. To fit the robot's navigation strategy, where These are the network parameters. After a certain number of training rounds, the network can complete the navigation task with a 100% success rate, indicating that the network has learned the task features necessary for navigation, although these features cannot be easily interpreted at present.

[0031] Furthermore, based on dictionary learning methods in large language models, an SAE layer of the MLP policy network is constructed to decompose the neural network superposition effect and explain the internal function of the MLP. Specifically, a layer such as... Figure 3 The sparse encoder and sparse decoder network shown uses SAE to recover a hidden layer in the MLP policy network. This can be formalized as:

[0032]

[0033] in The values ​​of the hidden layer of the original MLP network are used as input to the SAE module. Yes The estimated value of the reconstruction is used as the output of the SAE module. , These are the linear mapping parameters to be learned by the sparse encoder, and the JumpReLU function is the activation function used in SAE. For the sparse representation of SAE, it will ensure the sparsity of the representation through a norm-constrained loss. The sparse representation will then be decoded by a sparse decoder. The reconstruction here only uses matrix-vector multiplication; in this sense, the reconstruction is merely a... Each column A linear combination, therefore Also known as a feature dictionary, sparse representation It is equivalent to its coefficient.

[0034] The training of SAE consists of two parts: reconstruction loss and sparse loss. Reconstruction loss ensures that SAE is functionally aligned with the original MLP, while sparse loss ensures that the sparse representation z is simple, thus facilitating the acquisition of simple representation patterns. The loss function can be formalized as:

[0035] in The sparsity coefficients are used for this training method. This method is universal and therefore SAE can be applied to various current neural network models.

[0036] S2, Feature extraction and interpretation of sparse neurons in the strategy model.

[0037] A trained SAE can decompose the superposition of neurons in the policy model into sparse neurons, making the meaning of each neuron more singular. At this point, filtering can be used to select the effective features of the neural network and interpret them to a certain extent.

[0038] Analyzing the SAE trained in step 1 along the robot's trajectory reveals a clear correlation between the sparsity of the SAE and the robot's motion patterns. Figure 4 and Figure 5The data clearly illustrates these features. Due to the use of SAE, the activated features are in the minority. Furthermore, because their motion patterns are similar, their sparse representations are also similar. Comparing these two motion patterns reveals that features with horizontal axes of 2 and 61 are highly correlated with reversing and turning behavior, while the feature with horizontal axis of 26 is highly correlated with upward movement. Following a similar approach, a filtering algorithm using dictionary activation as a threshold was designed to filter features related to motion patterns. Specifically, for a class of similar motion data that has been filtered... Input all of them into the network and compute sparse representations. Filter by the product of the sparse representation and the dictionary. If the following conditions are met: Then it is considered that the dictionary Related to this movement pattern, This is the dictionary activation threshold; there may be multiple related dictionaries.

[0039] Therefore, for this robot task, patterns such as moving to the upper right, moving to the upper left, moving upwards, reversing to the left and turning around, and reversing to the right can be selected. Feature filtering based on these patterns can calculate the SAE sparse features of various movements.

[0040] S3 verifies interpretability by activating the patch intervention strategy model.

[0041] To verify the significance and correctness of the extracted features, activation patching can be used to intervene in the sparse activation values ​​of the SAE, thereby affecting the output values ​​of the neural network and thus changing the robot's motion strategy.

[0042] Specifically, based on the dictionary features selected in step 2 The sparse features of the SAE are constructed as coefficients and then inserted into the original neural network. The output motion pattern of the neural network is then observed to see if it matches expectations. Because SAE is used, only a few dimensions need to be constructed, reducing the human effort required for feature construction. The magnitude of the coefficient construction can be referenced from the results in the motion data.

[0043] As an example, the 26-dimensional feature is set to 0.5, and the other dimensions are set to 0. See the running graph. Figure 6 As can be seen, the robot can perform generally upward movements under continuous operation. However, because the values ​​are static and unchanging, it cannot perfectly stop within the target area. A more robust activation patch is expected to improve this.

[0044] The following specific embodiments further illustrate the robot motion strategy interpretability identification method based on sparse autoencoders of this application.

[0045] like Figure 2As shown, a simulation environment for a two-wheeled differential robot was rendered using pygame. The trajectory line changes from light red to dark red to indicate the timing of the motion, the triangular arrow indicates the direction of the vehicle's movement, and the pink square area represents the target area. The mission is considered successful when the vehicle enters the red area.

[0046] like Figure 3 As shown, the neural network acts as an intelligent agent for the robot, reading the state from the simulation environment, obtaining the robot's actions, and then executing them within the environment. A sparse autoencoder is inserted into one of the hidden layers to recover the hidden layer using a sparse representation. This sparse representation expresses the information contained in the hidden layer and can therefore serve as a low-dimensional representation of the network's function.

[0047] like Figure 4 As shown, in the left image, the robot's trajectory is represented by a red broken line, and the robot's orientation is represented by a blue arrow. The robot first reversed and turned, and then moved forward towards the target area. In the right image, the vertical axis represents the time scale of the trajectory, and the horizontal axis represents the dimension of sparse features. The lighter the color of the grid in the image, the stronger the activation of that coefficient feature at that time.

[0048] like Figure 5 As shown, the selection was made with Figure 4 Initial positions that are close in location but different angles result in similar but not identical subsequent motions, but the resulting sparse features are also similar.

[0049] like Figure 6 As shown, for reference Figure 4 and Figure 5 The activation status, given an approximate activation patch for upward movement, indicates the robot's trajectory.

[0050] This disclosure also provides a robot motion strategy interpretability identification device based on a sparse autoencoder, including: The training module is used to train the motion strategy of the mobile robot and the Sparse Autoencoder (SAE) layer: Two-wheel differential robot motion is used as a typical robot navigation task, and the imitation learning model trained on this task is used as a typical robot policy model. The SAE layer is then trained to analyze the policy model, which serves as a typical paradigm for policy model analysis. Feature extraction and interpretation, used for feature extraction and interpretation of sparse neurons in policy models: using filtering to select effective features of the neural network and interpreting them; The verification module is used to verify interpretability by activating the patch intervention strategy model.

[0051] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0052] In summary, this invention selects the motion of a two-wheeled differential speed robot as a representative robot task to study the interpretability issues in its neural network motion strategy. First, a robot simulation is built, and a teaching strategy is developed to collect data. A fully connected learning pipeline (MLP) network is used to train the motion strategy neural network through imitation learning. Then, an SAE layer is trained on this neural network to decouple the superposition effect in the neural network, facilitating analysis. Next, the SAE layer is analyzed to identify features representing typical motions and interpret their meaning. Finally, to verify the effectiveness of the motion features, their function is confirmed by activating patch intervention.

[0053] The method of this invention has the following advantages compared with existing research: (1) The superposition effect in the neural network is decoupled by using SAE layer, which makes the features of neurons more ambiguous and facilitates the analysis of the correlation between neuron patterns and motion strategies; (2) By activating the patch intervention, the effectiveness of the model interpretation was verified and the transparency of the neural network strategy was improved.

[0054] It should be noted that although several modules of the system for executing actions are mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more modules described above can be embodied in one module. Conversely, the features and functions of one module described above can be further divided into multiple modules for embodiment. Components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without any inventive effort.

[0055] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for robot motion policy interpretability identification based on sparse autoencoder, characterized in that, Comprising: S1, training of mobile robot motion strategy and training of sparse autoencoder SAE layer: using two-wheel differential robot motion as a typical robot navigation task, and using the imitation learning model trained on the task as a typical robot strategy model, and then training the SAE layer to analyze the strategy model, as a typical paradigm for strategy model analysis; S2, sparse neuron feature extraction and explanation in the strategy model: using filtering to screen out effective features of the neural network and explaining them; S3, verifying interpretability by activating patches to intervene in the strategy model.

2. The method of claim 1, wherein the method further comprises: In S1, the robot strategy model is an MLP network model with an input dimension of 9, a hidden layer of 64, and an output dimension of 2. 3.The method of claim 2, wherein, The MLP network model includes a sparse encoder and a sparse decoder network, which are represented as follows: in, The values ​​of the hidden layer of the original MLP network are used as input to the SAE module. Yes The estimated value of the reconstruction is used as the output of the SAE module. , These are the linear mapping parameters to be learned by the sparse encoder, and the JumpReLU function is the activation function used in SAE. This represents the sparse characterization of SAE. It is the parameter matrix of the sparse decoder in SAE (Sparse Autoencoder). Column vectors. 4.The method of claim 1, wherein, In S2, filtering is performed according to the product of sparse representation and dictionary.

5. The method of claim 4, wherein the method further comprises: If the following formula is satisfied: then the dictionary associated with the motion pattern, is the dictionary activation threshold, there can be multiple dictionaries associated.

6. The method of claim 5, wherein the method further comprises: With dictionary features The sparse features of SAE are constructed as coefficients, and the original neural network is intervened to observe whether the output motion pattern of the neural network meets the expectation.

7. A robot motion policy explainability recognition apparatus based on a sparse autoencoder, characterized by, Comprising: a training module for training of mobile robot motion strategy and training of sparse autoencoder SAE layer: using two-wheel differential robot motion as a typical robot navigation task, and using the imitation learning model trained on the task as a typical robot strategy model, and then training the SAE layer to analyze the strategy model, as a typical paradigm for strategy model analysis; a feature extraction and explanation module for sparse neuron feature extraction and explanation in the strategy model: using filtering to screen out effective features of the neural network and explaining them; a verification module for verifying interpretability by activating patches to intervene in the strategy model.

Images