A multi-task oriented behavior analysis method for rat robots

Abstract

The present invention discloses a multi-task oriented rat robot behavior analysis method, including a behavior recognition task module and a motion quality assessment task module. The behavior recognition task includes an input layer, a Transformer encoding module, a basic network module, a fully connected layer, and an output layer. The performance of behavior recognition is improved by placing the Transformer encoding module in front of the basic network module. The motion quality assessment task is implemented based on global and local fusion features, including low-dimensional mapping measurement, posture difference measurement based on quasi-GW distance, and body shaking degree measurement. By simultaneously measuring the high-level semantic information and internal time structure of the rat robot's key point information, the features are fused to accurately judge the motion smoothness. Based on the key point information, the present invention can simultaneously achieve the dual tasks of behavior classification and motion quality assessment of the rat robot, with high classification accuracy and strong interpretability of the motion quality assessment.

Description

Technical Field

[0001] The present invention belongs to the fields of motion quality assessment technology, Transformer technology, and behavior recognition technology, and particularly relates to a multi-task oriented rat robot behavior analysis method. Background Art

[0002] Rat robots are a new type of biorobot with broad application prospects in fields such as biomedical research, military reconnaissance, and rehabilitation therapy. The quality of their motion directly impacts the effectiveness and performance of rat robots in practical applications. In applications, when sequence control and state transitions are not smooth, the movements of rat robots can become clumsy and dangerous, leading to low efficiency and even mission failure. Therefore, evaluating the motion quality of rat robots is essential. This can help researchers understand the capabilities and limitations of rat robots and provide a scientific basis for their further improvement and optimization.

[0003] Existing motion quality assessment methods have limitations in both qualitative and quantitative descriptions. For example, they only consider action recognition results, which makes it difficult to further improve accuracy, and the assessment methods lack interpretability.

[0004] Specifically, when it comes to assessing the quality of rat robot motion, there's a lack of clear standards for defining smoothness. While defining "smoothness" based solely on one-dimensional data is relatively straightforward, in reality, rat robot motion has multiple attributes, such as speed and posture. It's possible for a rat robot to be smooth in one attribute while abnormal in another. Therefore, defining a "smoothness" range that covers all possible normal behaviors is extremely difficult. A second challenge is that the boundary between smooth and abnormal motion is often imprecise; abnormal behaviors near the boundary may actually be smooth, and vice versa. A third challenge is the imbalanced distribution of rat motion data, with a larger proportion of natural motion and fewer abnormal movements. This is because natural motion is easier to obtain, while abnormal data such as spasms and gallops is more difficult to obtain. Therefore, assessing motion smoothness and identifying abnormal behaviors from a limited set of abnormal behavior sequences is a challenging task.

[0005] Behavior recognition is a key research area in computer vision and pattern recognition. Compared to human behavior recognition, animal behavior recognition faces more challenges, such as significant behavioral variations among different animal species, limited camera field of view, and numerous influencing factors such as lighting and weather conditions. Animal behavior plays a key role in the intelligent control of animal robots and in assisting in establishing the correspondence between specific animal behaviors and nervous system activity. Therefore, accurately identifying animal behavior is a crucial research topic. Traditional action recognition methods first use a clustering algorithm to cluster 3D skeleton points into k pose sequences. These pose sequences are then subjected to behavior recognition using an HMM model. Mainstream deep learning methods include two-stream, 3D convolution (C3D), CNN+RNN, and graph convolution. The two-stream method assumes that video is composed of consecutive frames, thus being continuous in time. Each frame contains spatial information, recording the object's motion and scene information. Therefore, video data possesses spatiotemporal characteristics.

[0006] The two-stream network uses two branches to extract spatiotemporal features to handle action recognition tasks. One branch, the temporal stream, extracts motion features through dense optical flow between adjacent frames. The other branch, the spatial stream, extracts object and scene appearance features from an RGB image extracted from the input video. The two feature maps are fused, and the output layer ultimately outputs the action results. However, the accuracy of the two-stream network decreases significantly when analyzing longer videos.

[0007] The C3D method believes that three-dimensional convolutional neural networks can not only learn spatial features, but also correspond to the time dimension in videos due to the additional dimension compared to 2D. Three-dimensional convolutional neural networks do not require other preprocessing steps when processing videos, so they are universally applicable to video classification tasks. Compared with the two-stream method, the C3D method does not use the operation of separating and then fusing spatiotemporal features, so it can capture spatiotemporal features at the same time. However, since the number of parameters of three-dimensional convolutional neural networks is much larger than that of two-dimensional convolutional neural networks, not only is overfitting prone to occur on small amounts of data during training, but it also increases the overall training time.

[0008] Because behaviors are continuous in time, a new behavior recognition architecture combining CNNs and LSTMs was considered. This modeling approach leverages the visual feature extraction strengths of CNNs and the temporal modeling strengths of RNNs, achieving promising results in behavior recognition tasks. Currently, there are two main strategies for keypoint-based behavior recognition: directly utilizing the underlying model to build a classification network, and utilizing keypoints to construct a graphical model for spatiotemporal learning to derive behavior categories. However, directly utilizing the underlying model does not achieve the required accuracy, while constructing a graphical model requires redesigning the skeletal graph structure when the number of research objects or keypoints changes, preventing reuse of the existing graph structure. Therefore, keypoint-based behavior recognition still presents significant challenges. Summary of the Invention

[0009] The purpose of the present invention is to propose a multi-task oriented rat robot behavior analysis method to solve the multi-task problem of behavior recognition and motion quality assessment.

[0010] In order to solve the above technical problems, the present invention specifically provides the following technical solutions: a multi-task oriented rat robot behavior analysis method, which mainly includes a behavior recognition task module and a motion quality assessment task module.

[0011] The behavior recognition task is implemented based on the Transformer encoding module, including an input layer, a Transformer encoding module, a basic network module, a fully connected layer, and an output layer. Based on the key point sequence obtained by the posture estimation method, the performance of behavior recognition is improved by prepending the Transformer encoding module to the basic network module.

[0012] The motion quality assessment task is implemented based on a fusion of global and local features, including a low-dimensional mapping metric, a posture difference metric based on quasi-GW distance, and a body jitter metric. The low-dimensional mapping metric generates global feature information, while the posture difference metric based on quasi-GW distance and the body jitter metric each generate local feature information. Ultimately, the fused feature metrics yield a motion smoothness score.

[0013] The behavior recognition model based on the Transformer encoding module has a serial structure. The order of the modules is from first to last: input layer, Transformer encoding module, basic network module, fully connected layer, and output layer. The basic network module can adopt a variety of neural network structures, such as one-dimensional convolutional neural network (1DCNN), multi-layer perceptron (MLP), recurrent neural network (RNN), long short-term memory network (LSTM), gated recurrent unit (GRU), bidirectional long short-term memory network (BiLSTM), bidirectional gated recurrent unit (BiGRU) and other methods.

[0014] The overall data processing flow for the behavior recognition task includes: first, continuous frame data is input into the input layer, features are extracted by the Transformer encoding module, and then sent to the base network module. Feature filtering is then performed in the fully connected layer, and finally the classification result is generated by the output layer. The input layer is used to receive raw data, which is the coordinate point information of the rat robot. This raw data can be one-dimensional x / y / z coordinate point information, two-dimensional x,y / y,z / x,z coordinate point information, or three-dimensional x,y,z coordinate point information. The Transformer encoding module is used to receive and process the data transmitted by the input layer, extracting spatiotemporal features from the raw data through feature encoding, which can be efficiently learned by the neural network. The base network module is the learning part of the model architecture, capable of learning the features processed by the Transformer encoding module and further extracting higher-level abstract features to better support the completion of the behavior recognition task. The fully connected layer functions to connect each node output by the base network module to all nodes in the previous layer, integrating the features extracted by the base network module and mapping the learned "distributed feature representation" to the sample label space. The output layer performs the final classification on the feature data output by the fully connected layer to obtain the final behavior classification result.

[0015] A multi-task rat robot behavior analysis method is proposed. The motion quality assessment task is implemented based on a fusion of global and local features. The method comprises an input layer, a low-dimensional mapping metric, a posture difference metric based on quasi-GW distance, a body jitter metric, and an output layer. The input layer receives the coordinate point information of the rat robot. The low-dimensional mapping metric obtains global features, uses dimensionality reduction techniques to map high-dimensional sampled data into a low-dimensional space, and calculates the probability of the tested sequence falling outside the smooth motion range. The posture difference metric based on quasi-GW distance obtains local motion features and establishes a difference matrix of all key point posture data of the rat robot to measure the differences between adjacent postures. The body jitter metric uses spectral analysis to calculate frequency components, mapping jitter changes that are difficult to describe mathematically in the time domain onto a spectral curve. Energy distribution characteristics are used to determine whether abnormal motion has occurred. The output layer outputs the rat robot's motion smoothness score.

[0016] The global features and local features used in the motion quality assessment task are complementary. After "scoring", normalizing and assigning weights to each dimension of features respectively, the features are linearly fused to obtain the motion smoothness score, which can be used to effectively determine whether the motion sequence is abnormal.

[0017] Another object of the present invention is to provide the application of the rat robot behavior analysis method for behavioral classification and motion quality assessment. Through detailed analysis of rat behavior, this method can better understand its behavioral characteristics and neurobiological mechanisms, providing valuable information for neuroscience research, fluid control, rehabilitation therapy, and other fields. It can also be applied in biomedical research, military reconnaissance, and disaster rescue.

[0018] The method of the present invention can be applied to biomedical research, military reconnaissance and disaster rescue. Compared with traditional robots, rat robots use their bodies directly as the main body, without the need to consider material manufacturing, and retain the good functions of the animals themselves through natural evolution. Rat robots can use visual, chemical, mechanical and magnetic receptors to perceive external stimuli, and can use existing capabilities to independently achieve motion control and obstacle avoidance functions in complex environments. Rat robots have strong environmental adaptability, flexible movements, strong concealment, and low power consumption. In the future, they will have broad application prospects in the fields of industry, military, life, and medical care. They can be used for disaster rescue, military reconnaissance, navigation, and dangerous goods inspection. At the same time, we can obtain new and even unknown information from the direct interaction between rats and the environment.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] (1) Based on the rat robot posture estimation method, the present invention proposes a multi-task solution to achieve qualitative and quantitative behavior detection, which can complete behavior recognition and generate motion quality assessment scores.

[0021] (2) In terms of motion quality assessment, the present invention addresses the problems of lack of clear standards and lack of interpretability in motion quality assessment methods. It linearly fuses global features and local features, and combines multiple different types of kinematic properties to improve the accuracy of rat robot motion smoothness assessment, which has good interpretability.

[0022] (3) In terms of behavior recognition, the present invention addresses the universality problem that the behavior recognition method based on key points needs to manually redesign the skeleton graph structure according to the research object, and the problem that the behavior recognition method based on key points has low accuracy in directly classifying temporal key points. By utilizing the feature extraction capability of the Transformer encoding module, the present invention can not only automatically design the optimal skeleton graph structure directly according to the key point information, avoiding the manual design of the skeleton graph structure, and having higher universality, but also greatly improve the accuracy of the basic behavior recognition model. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention, the drawings required for describing the embodiments are briefly introduced below.

[0024] Figure 1This is a framework diagram of the multi-task rat robot behavior analysis method of the present invention.

[0025] Figure 2 This is a schematic diagram of the Transformer encoding module structure of the present invention.

[0026] Figure 3 This is an example diagram of an action recognition model based on the Transformer encoding module. DETAILED DESCRIPTION

[0027] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0028] Example 1

[0029] like Figure 1 As shown, the present invention provides a multi-task oriented rat robot behavior analysis method, which is used to evaluate the motion quality of the rat robot based on the posture estimation method, and mainly includes a behavior recognition task module and a motion quality assessment task module. The behavior recognition task is implemented based on the Transformer encoding module, including an input layer, a Transformer encoding module, a basic network module, a fully connected layer, and an output layer. The behavior recognition method based on key points improves the performance of behavior recognition by prepending the Transformer encoding module to the basic network module. The motion quality assessment task is implemented based on global and local fusion features, including low-dimensional mapping metrics, posture difference metrics based on quasi-GW distance, and body jitter degree metrics. The motion smoothness score is obtained by fusion feature metrics.

[0030] A multi-task oriented rat robot behavior analysis method, wherein a behavior recognition model based on a Transformer encoding module has a serial structure, and the order of the modules is, from first to last, an input layer, a Transformer encoding module, a basic network module, a fully connected layer, and an output layer. The basic network module can adopt a variety of neural network structures, such as a one-dimensional convolutional neural network (1DCNN), a multi-layer perceptron (MLP), a recurrent neural network (RNN), a long short-term memory network (LSTM), a gated recurrent unit (GRU), a bidirectional long short-term memory network (BiLSTM), a bidirectional gated recurrent unit (BiGRU), and the like.

[0031] The overall data processing flow for the behavior recognition task involves first inputting continuous frame data into the input layer, extracting features through the Transformer encoding module, then feeding it into the base network module. Feature filtering is then performed in the fully connected layer, and finally, the output layer generates the classification result. The input layer receives raw data, which is the coordinate point information of the rat robot. This data can be one-dimensional x / y / z coordinate point information, two-dimensional x,y / y,z / x,z coordinate point information, or three-dimensional x,y,z coordinate point information. The Transformer encoding module receives and processes the data passed by the input layer, extracting spatiotemporal features from the raw data through feature encoding, which can be efficiently learned by the neural network. The base network module is the learning component of the model architecture, capable of learning the features processed by the Transformer encoding module and further extracting higher-level abstract features to better support the completion of the behavior recognition task. The fully connected layer connects each node output by the base network module to all nodes in the previous layer, integrating the features extracted by the base network module and mapping the learned "distributed feature representation" to the sample label space. The output layer performs the final classification on the feature data output by the fully connected layer to obtain the final behavior classification result.

[0032] like Figure 2As shown in Figure 1, the Transformer encoding module includes input embedding, positional encoding, multi-head self-attention, normalization, feedforward, and residual connections. Input embedding maps the input sequence into a low-dimensional vector for easier neural network processing. Positional encoding adds position information to each input vector. These positional encoding vectors are added to the input embedding vector to incorporate the position information into the input embedding. Each sub-attention head of the multi-head self-attention mechanism performs self-attention on the vectors in the input sequence, learning similarities between different vectors and improving the model's representational power. Normalization standardizes the input data to ensure that it is distributed within an appropriate range, thereby improving the stability and learning performance of the neural network. Feedforward, consisting of a nonlinear activation function and a linear transformation, is applied to the output of each sub-layer. A fully connected layer maps and transforms the sub-layer output, further improving the model's representational power. Residual connections are applied between the input and output of each sub-layer to ensure stability during model training. Transformer can be considered a graph modeling method. Its graph is fully connected, and the relationships between nodes are learned in a data-driven manner. Because any concrete or abstract concept can be represented by a node in the graph, and the relationships between concepts can be characterized by edges on the graph, Transformer modeling is highly versatile.

[0033] like Figure 3 As shown, the basic structure of the behavior recognition model based on the Transformer encoding module is composed of an input layer, a Transformer encoding module, a bidirectional recurrent neural network (BiGRU), a Pooling layer, a Dropout layer, a Dense layer, etc. Among them, the Dense layer is used as the final classifier. Since behavior is an action that continues in time, its long-term dependencies need to be considered. The bidirectional recurrent neural network (BiGRU) can connect the output of any node with the state of the context, so it can better capture the long-term dependencies of the action. This implementation example uses the general graph modeling capability of the Transformer encoding module to first perform feature encoding on the three-dimensional posture data. The input dimension of the dataset is 30×18, where the first dimension 30 refers to the length of the behavior, which can be changed for different behavior standards; when the number of key points is 6, and each point has three coordinate values of x, y, and z, the second dimension value is 18.

[0034] A multi-task rat robot behavior analysis method is proposed. The motion quality assessment task is implemented based on a fusion of global and local features. The method comprises an input layer, a low-dimensional mapping metric, a posture difference metric based on quasi-GW distance, a body jitter metric, and an output layer. The input layer receives the coordinate point information of the rat robot. The low-dimensional mapping metric obtains global motion features, maps high-dimensional sampled data into a low-dimensional space using dimensionality reduction methods, and calculates the probability of the tested sequence falling outside the smooth motion range. The posture difference metric based on quasi-GW distance obtains local motion features and measures the differences between adjacent postures by establishing a difference matrix for all key point posture data of the rat robot. The body jitter metric obtains local motion features and uses spectral analysis methods to calculate frequency components. Jitter changes that are difficult to describe mathematically in the time domain are mapped onto a spectral curve, and energy distribution characteristics are used to determine whether abnormal motion has occurred. The output layer outputs the rat robot's motion smoothness score.

[0035] The low-dimensional mapping metric is to build a model from the normal sequence to determine the range of smooth natural motion and calculate the probability of the sequence to be tested outside the ideal range. Smooth motion and abnormal motion are essentially separable and there are significant differences. The principal component analysis (PCA) technique is used to map the high-dimensional sampled data into a low-dimensional space. For N normal samples x1~x N , each sample dimension is P×V×2, P represents the number of frames, V represents the number of key points, and the normal sample matrix X is defined as:

[0036]

[0037] The distance between the reduced-dimensional sample S and the mean of all normal sample matrices is calculated in the feature space and expressed as:

[0038] Here, s represents the value of the sample S to be tested in feature space. Smooth and abnormal motion exhibit separate concentric annular distributions in feature space, with the difference between the two distributions reflected in the distance from the center of mass. Smooth motion, defined as rat motion without any electrical stimulation, exhibits a nonlinear annular distribution in feature space, moving away from the center of mass, while abnormal motion is distributed near the center of mass. The closer the data is to the center of mass, the higher the degree of abnormality.

[0039] The pose difference metric based on the quasi-GW distance is equivalent to comparing the degree of pose change between two adjacent time steps. Specifically, a skeleton sequence is constructed using the key points of the rat robot. The individual rat is abstracted into a two-dimensional fine skeleton representation described by the pose key points. The quasi-GW distance, which is suitable for non-rigid deformation and has good stability, is used to analyze the correspondence between any two key points in the fine pose. A difference matrix for all the key point pose data of the rat robot is constructed, thus obtaining a comprehensive difference metric between the two sets of skeletons.

[0040] For a pose variable X∈R with T poses T×V×2 , each posture has V key points, and the two-dimensional coordinates of each key point are (x, y), then for any two postures p,q∈R V×2 , calculate the quasi-GW distance between two sets of pose sequences under a fixed transmission scheme:

[0041] c(p,q)=||M p -M q ||

[0042] Where M represents the pairwise distance matrix, M i,j =||p i -p j If the pose differences are large, the quasi-GW distance will be larger. By calculating the pose differences between two adjacent frames, we can determine when significant pose switching occurs in the pose sequence, promptly identifying segments that differ significantly from natural motion. By adjusting the control strategy, we can achieve smooth transitions and sequential control between these motion states.

[0043] The body shaking degree measurement is used as the basis for the degree of smoothness to measure the degree of spasmodic shaking of the body. When the rat robot moves abnormally, the limbs are not coordinated and the body shakes obviously. This shaking will cause the energy of certain frequency bands of the speed to change. The spectrum analysis method can calculate the energy of each frequency band. Since the normal and abnormal frequency components are different, it is possible to judge whether abnormal movement occurs by the energy distribution characteristics, and map the shaking changes that are difficult to describe mathematically in the time domain onto the spectrum curve. After using the fast Fourier transform (FFT) method to transform the speed characteristics from the time domain to the frequency domain, the comparison of the normal and abnormal movement speed spectrum curves can effectively identify the body shaking degree of the energy distribution during the movement of the rat robot. The larger the area of the rat shaking frequency band range, the greater the potential abnormality of the rat robot.

[0044] We use global and local features to compute kinematic features from the motion capture data for the task. We then "score," normalize, and weight each dimension of the feature. Inspired by existing model-based fusion methods, we fuse the feature values to ultimately place the overall motion smoothness in the range [0, 1], where 0 represents idealized motion quality and 1 represents the maximum deviation from idealized motion. The smoothness evaluation function y = f(x) is a linear combination of the kinematic properties:

[0045] is to convert a single kinematic variable x i Normalized continuous function, ω i is the weight of this variable. In the motion quality assessment task, the key point sequence on the rat robot is input, and after weighted fusion of global and local features, the output is the motion smoothness score.

[0046] Through the multi-task rat robot behavior analysis method of this embodiment, the dual tasks of behavior classification and movement fluency evaluation of the rat robot can be simultaneously achieved based on key point information, with high accuracy of behavior classification and strong interpretability of movement quality evaluation.

[0047] The above embodiments are merely exemplary embodiments of the present application and are not intended to limit the scope of the present application. The scope of protection of the present application is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present application within the essence and scope of protection of the present application, and such modifications or equivalent substitutions shall also be deemed to fall within the scope of protection of the present application.

Claims

1. A multi-task oriented rat robot behavior analysis method, characterized in that: It includes a behavior recognition task module and a motion quality assessment task module. The behavior recognition task is implemented based on the Transformer encoding module, including the input layer, Transformer encoding module, basic network module, fully connected layer, and output layer. Based on the key point sequence obtained by the posture estimation method, the performance of behavior recognition is improved by pre-placing the Transformer encoding module in the basic network module. The motion quality assessment task is implemented based on global and local fusion features, including low-dimensional mapping measurement, posture difference measurement based on quasi-GW distance, and body jitter degree measurement. The motion smoothness score is obtained by fusion feature measurement. The overall data processing flow for the behavior recognition task includes: first, continuous frame data is input into the input layer, features are extracted through the Transformer encoding module, and then sent to the basic network module, followed by feature screening through the fully connected layer, and finally the classification result is generated by the output layer; the input layer is used to receive the original data, and the original data is the coordinate point information of the rat robot, which can be one-dimensional x / y / z coordinate point information, two-dimensional x, y / y, z / x, z coordinate point information, or three-dimensional x, y, z coordinate point information, the Transformer encoding module is used to receive and process the data passed by the input layer, and extract features through feature encoding. The spatiotemporal features in the raw data can be efficiently learned by the neural network. The basic network module is the learning part of the model architecture, which can learn the features processed by the Transformer encoding module and further extract higher-level abstract features to better support the completion of the behavior recognition task. The function of the fully connected layer is to connect each node output by the basic network module with all nodes in the previous layer, to integrate the features extracted by the basic network module and map the learned "distributed feature representation" to the sample label space. The output layer is to perform the final classification of the feature data output by the fully connected layer to obtain the final behavior classification result. Descriptors of global and local feature fields are created to accurately judge the smoothness of movement by simultaneously measuring the high-level semantic information and internal temporal structure of the rat robot's key point information. The global and local features used for the motion quality assessment task are complementary. After "scoring", normalizing and weighting each dimension of the feature, the features are linearly fused to obtain the motion smoothness score, which can be used to effectively determine whether the movement sequence is abnormal.

2. A multi-task oriented rat robot behavior analysis method according to claim 1, characterized in that: The behavior recognition model based on the Transformer encoding module has a serial structure. The order of the modules is, from first to last, the input layer, the Transformer encoding module, the basic network module, the fully connected layer, and the output layer. The basic network module adopts a variety of neural network structures, and selects one of the following methods: one-dimensional convolutional neural network, multi-layer perceptron, recurrent neural network, long short-term memory network, gated recurrent unit, bidirectional long short-term memory network or bidirectional gated recurrent unit.

3. A multi-task oriented rat robot behavior analysis method according to claim 1, characterized in that: The motion quality assessment model based on the fusion of global and local features includes: an input layer, a low-dimensional mapping metric, a posture difference metric based on quasi-GW distance, a body jitter degree metric, and an output layer; wherein, the input layer is used to receive the coordinate point information of the rat robot, the low-dimensional mapping metric obtains global features, uses dimensionality reduction technology to map high-dimensional sampling data to a low-dimensional space, and calculates the possibility of the sequence to be tested outside the smooth motion range; the posture difference metric based on quasi-GW distance obtains local features, and by establishing a difference matrix of all key point posture data of the rat robot, it is used to measure the difference between adjacent postures; the body jitter degree metric obtains local features, uses a spectrum analysis method to calculate the frequency component, maps the jitter changes that are difficult to describe mathematically in the time domain onto the spectrum curve, and determines whether abnormal motion occurs through energy distribution characteristics; the output layer outputs the motion smoothness score of the rat robot.

4. A rat robot behavior classification and motion quality assessment system, which implements the multi-task oriented rat robot behavior analysis method as claimed in claim 1.

Images