A humanoid robot gait analysis and recognition and robot control method

Abstract

The present invention discloses a humanoid robot gait analysis and robot control method, and relates to the field of robot control technology. The collected gait data is used to pre-train a gait analysis and recognition model, which is used to evaluate the gait parameters of the humanoid robot. The gait parameters include: gait symmetry, gait consistency, movement, gait phase, and gait health; gait parameters estimated in real time based on the pre-trained gait analysis and recognition model. Afterwards, the pre-trained gait analysis and recognition model is deployed into the humanoid robot control strategy; the gait parameters estimated in real time by the gait analysis and recognition model are added to the state space of the humanoid robot control strategy, and the control strategy generates and executes corresponding behavioral actions. The present invention evaluates the symmetry and consistency of the humanoid robot during movement, and simultaneously identifies the robot's movements and gait phases to quantify the robot's gait parameters, so as to improve the robot's performance during the learning process.

Description

Technical Field

[0001] The present invention relates to the field of robot control technology, and in particular to a humanoid robot gait analysis and recognition method based on a pre-training model and a robot control method. Background Art

[0002] Humanoid robots can adapt to indoor environments and use tools designed for humans, and have broad application prospects in home services. However, the gait control of humanoid robots is still in its infancy, and there are many difficulties that need to be solved. First, most of the existing gait control strategies output the joint movements to be performed by processing raw sensor data. This method does not allow the robot to learn high-level gait rules. Secondly, robot control is dangerous and unknown. A human-sized robot that loses balance and falls can cause very serious injuries. Therefore, real-time monitoring of the robot's status is a very important process. However, there is currently no unified standard in the field of robotics to analyze gait.

[0003] Gait symmetry, consistency, and health status can be used as important parameters to help the robot adjust to a reasonable gait, and the estimated robot motion can provide prior information for the next step of robot control. The five humanoid robot gait evaluation indicators proposed in this invention are introduced as follows.

[0004] 1) Gait symmetry: Symmetry is an important indicator for evaluating human gait, and it also has great research value in the field of robotics. Gait symmetry can be used to evaluate the stability and movement standardization of robots. By analyzing the symmetry of robots when walking or running, it can also help them adjust their posture and improve their movement efficiency. In addition, the asymmetry of the robot's gait may also increase energy consumption. Therefore, the study of gait symmetry is of great significance for improving the robot's movement performance and bio-simulation capabilities.

[0005] 2) Gait consistency: Gait consistency is an important indicator for evaluating the stability and reliability of a robot's gait. It can reflect the stability of a robot's gait under the same conditions. A robot with good gait consistency can maintain a stable gait when performing the same task, and has higher reliability and safety. This is especially important for robots that need to run for a long time or perform tasks in a specific environment. In the present invention, the gait consistency indicates the degree of change in the gait symmetry of the robot under multiple movements under the same conditions. The smaller the degree of change, the better the gait consistency.

[0006] 3) Gait cycle status: Gait is the posture of the human body during movement, which is periodic and regular. A complete gait cycle starts with the initial contact of the foot with the ground and ends with the heel of the same side touching the ground again. The gait cycle can be mainly divided into two stages: the support phase and the swing phase. The support phase refers to the stage when the foot maintains contact with the ground, accounting for 62% of the entire gait cycle; the swing phase refers to the stage when the foot swings in the air without contact with the ground, accounting for 38% of the entire gait cycle. By dividing and analyzing the gait cycle parameters, the posture changes during movement can be characterized. Gait phase is an important indicator reflecting gait habits and body coordination, and has important application value in the field of robot gait generation.

[0007] 4) Action recognition: Human activity recognition is an active research field that aims to identify human actions and behaviors by analyzing sensor data or image data. In the field of robotics, action recognition of robots can improve the intelligence level of robots, enabling them to interact and cooperate with the environment more flexibly and efficiently. The actions that can be recognized by the present invention include standing, sitting, lying, walking, running, going up and down slopes, and going up and down stairs.

[0008] 5) Health detection: Gait-based human health detection is a non-invasive, simple and effective method. Through gait analysis, important information about the primary motor cortex, basal ganglia and cerebellum functions of the human body can be extracted, which can provide important references for health management and disease prevention. Similarly, by analyzing the gait of a humanoid robot, abnormal gait of the robot can be detected, helping the robot to adjust the control strategy in time to achieve more reasonable movement. Summary of the invention

[0009] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a humanoid robot gait analysis and robot control method based on a pre-trained model and a robot control method.

[0010] The purpose of the present invention can be achieved by the following technical solutions:

[0011] As a first aspect of the present invention, a method for analyzing gait of a humanoid robot is provided, the method comprising the steps of:

[0012] Using the collected gait data to pre-train a gait analysis and recognition model to evaluate the gait parameters of the humanoid robot, the gait parameters include: gait symmetry, gait consistency, movement, gait phase, and gait health;

[0013] Gait parameters estimated in real time based on a pre-trained gait analysis and recognition model.

[0014] As a preferred technical solution, the gait analysis and recognition model uses a dynamic time warping algorithm to calculate the waveform similarity of the time domain signals of the left and right legs to obtain gait symmetry, as follows:

[0015] For the left and right leg time series, an n×m matrix grid is constructed respectively. The position (i, j) of the matrix is ​​used to store point a i and point b j The Euclidean distance of

[0016] In the matrix grid, the continuous set of elements constitutes the path as a regular path W;

[0017] Find the path with the shortest cumulative distance:

[0018]

[0019] Among them, K represents the number of path points, ω k represents the kth element of the regularized path W;

[0020] Initialize the distance matrix D and set D(1,1) to the distance between the first element of the left leg time series A and the right leg time series B;

[0021] Starting from the starting point, the minimum accumulated value D(n,m) obtained by iterative calculation is the shortest accumulated distance DTW(A,B) between the left leg time series and the right leg time series.

[0022] As a preferred technical solution, the gait analysis and recognition model evaluates the consistency of gait by calculating the variance of gait symmetry obtained by multiple repeated movements.

[0023] As a preferred technical solution, the real-time action recognition of the gait analysis and recognition model adopts a ConvLSTM-based action classification model to classify actions. The action classification model is composed of a convolutional layer, an LSTM layer, a MaxPool2D layer and a fully connected layer, wherein the output layer of the action classification model is provided with a plurality of neurons for classifying and representing different types of actions.

[0024] As a preferred technical solution, the real-time action recognition adopts a sliding window mechanism to detect the robot's actions in real time, intercepts the gait signal based on the set window length, window movement frequency and overlap rate, and inputs the intercepted motion data into the ConvLSTM-based action classification model to recognize the action in real time.

[0025] As a preferred technical solution, the gait health assessment of the gait analysis and recognition model is specifically as follows:

[0026] The gaits of healthy subjects and patients are collected to construct a human gait dataset and train a health assessment model; the health assessment model adopts a ConvLSTM model, which is composed of a convolutional layer, an LSTM layer, a MaxPool2D layer and a fully connected layer, wherein the number of neurons in the output layer of the health assessment model is set to 1.

[0027] As a preferred technical solution, the classification of gait phases includes: heel strike, foot flat, mid-stance, heel lift, toe lift, swing cycle and non-walking movements in a complete gait cycle; the gait phases are evaluated using a fully connected network model.

[0028] As a second aspect of the present invention, a robot control method is provided, the method steps comprising:

[0029] Establishing a pre-trained gait analysis and recognition model based on any of the above humanoid robot gait analysis methods;

[0030] deploying the pre-trained gait analysis and recognition model into a humanoid robot control strategy;

[0031] The gait parameters estimated in real time by the gait analysis and recognition model are added to the state space of the humanoid robot control strategy, and the control strategy generates and executes corresponding behavioral actions.

[0032] As a preferred technical solution, the state space includes: the position and velocity of the humanoid robot's joints, the relative position, direction, linear velocity and angular velocity of the humanoid robot's joints relative to the root node in the local coordinate system, the height of the humanoid robot's pelvis from the ground, and the robot's gait symmetry, gait consistency, movement, gait phase and gait health.

[0033] As a preferred technical solution, the humanoid robot control strategy uses the proximal strategy optimization algorithm as the core framework of reinforcement learning, and adopts a fully connected network to construct a strategy network and a value network;

[0034] The strategy network is used to generate and execute corresponding behavioral actions, and the value network is responsible for evaluating the value brought by the actions generated by the strategy network.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] The present invention proposes a gait analysis and recognition algorithm for evaluating the symmetry and consistency of a humanoid robot during movement, and for identifying the robot's movements and gait phases to quantify the robot's gait parameters. Specifically, a gait analysis and recognition model is trained by collecting large-scale gait data of a human or humanoid robot, and this pre-trained model is deployed in a humanoid robot to evaluate a variety of gait parameters of the humanoid robot in real time. At the same time, the estimated gait parameters are added to the state space of a reinforcement learning algorithm to improve the robot's performance during the learning process. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a flow chart of a method for analyzing and identifying gait of a humanoid robot and controlling the robot according to the present invention;

[0038] Figure 2 This is a schematic diagram of the framework of the action recognition model of the present invention;

[0039] Figure 3 Set up a schematic for the sliding window. DETAILED DESCRIPTION

[0040] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0041] Gait symmetry, consistency, and health status can be used as important parameters to help the robot adjust to a reasonable gait, and the estimated robot motion can provide prior information for the next step of robot control. The five humanoid robot gait evaluation indicators proposed in this invention are introduced as follows.

[0042] 1. Gait symmetry

[0043] Symmetry is an important indicator for evaluating human gait, and it also has great research value in the field of robotics. Gait symmetry can be used to evaluate the stability and movement norms of robots. By analyzing the symmetry of robots when walking or running, it can also help them adjust their posture and improve their movement efficiency. In addition, the asymmetry of the robot's gait may also increase energy consumption. Therefore, the study of gait symmetry is of great significance for improving the robot's movement performance and bio-simulation capabilities.

[0044] 2. Gait consistency

[0045] Gait consistency is an important indicator for evaluating the stability and reliability of a robot's gait. It can reflect the stability of a robot's gait under the same conditions. A robot with good gait consistency can maintain a stable gait when performing the same task, and has higher reliability and safety. This is especially important for robots that need to run for a long time or perform tasks in a specific environment. In the present invention, the gait consistency indicates the degree of change in the gait symmetry of the robot under multiple movements under the same conditions. The smaller the degree of change, the better the gait consistency.

[0046] 3. Gait cycle status

[0047] Gait is the posture of the human body during movement. It is periodic and regular. A complete gait cycle starts from the initial contact of the foot with the ground and ends when the heel of the same side touches the ground again. The gait cycle can be mainly divided into two stages: the support phase and the swing phase. The support phase refers to the stage when the foot is in contact with the ground, accounting for 62% of the entire gait cycle; the swing phase refers to the stage when the foot swings in the air without contacting the ground, accounting for 38% of the entire gait cycle. By dividing and analyzing the parameters of the gait cycle, the posture changes during movement can be characterized. The gait phase is an important indicator reflecting gait habits and body coordination, and has important application value in the field of robot gait generation.

[0048] 4. Action Recognition

[0049] Human activity recognition is an active research field that aims to identify human movements and behaviors by analyzing sensor data or image data. In the field of robotics, motion recognition of robots can improve the intelligence level of robots, enabling them to interact and cooperate with the environment more flexibly and efficiently. The actions that can be recognized by the present invention include standing, sitting, lying, walking, running, going up and down slopes, and going up and down stairs.

[0050] 5. Health Check

[0051] Human health detection based on gait is a non-invasive, simple and effective method. Through gait analysis, important information about the primary motor cortex, basal ganglia and cerebellum functions of the human body can be extracted, which can provide important references for health management and disease prevention. Similarly, by analyzing the gait of a humanoid robot, abnormal gait of the robot can be detected, helping the robot to adjust the control strategy in time to achieve more reasonable movement.

[0052] Example 1

[0053] As one of the implementation modes of the present invention, this embodiment provides a humanoid robot gait analysis method based on a pre-trained model, which is used to evaluate the symmetry and consistency of the humanoid robot during movement, and identify the robot's movements and gait phases to quantify the robot's gait parameters. By collecting large-scale gait data of the human body or humanoid robot, a gait analysis and recognition model is trained, and this pre-trained model is deployed in the humanoid robot to evaluate a variety of gait parameters of the humanoid robot in real time. At the same time, the estimated gait parameters can also be added to the state space of the reinforcement learning algorithm to improve the robot's performance during the learning process. Figure 1 As shown, the method mainly includes the following steps:

[0054] Step 1: In the pre-training stage, a gait analysis and recognition model is trained using large-scale gait data collected from the human or humanoid robot field to evaluate the symmetry and consistency of the humanoid robot during movement, and to identify the robot's movements and gait phases to quantify the robot's gait parameters. The evaluation model of the proposed multiple gait indicators is designed as follows:

[0055] 1.1 Gait symmetry

[0056] Dynamic Time Warping (DTW) is a method for comparing the similarity between two time series, which can solve the matching problem between two sequences with different lengths on the time axis. The basic idea of ​​the DTW algorithm is to dynamically adjust the time axes of the two sequences so that they are aligned in the time dimension, thereby calculating the similarity between them. The present invention introduces the DTW algorithm to calculate the waveform similarity of the time domain signals of the left and right legs to obtain gait symmetry.

[0057] The following are the detailed steps to apply the DTW algorithm to calculate the waveform similarity of the left and right leg gait signals:

[0058] Assume that the time series of the left and right legs are A={a 1 ,a 2 ,...,a i ,...,a n},B={b 1 ,b 2 ,...,b j ,...,b m}, n is the length of the left leg time series, and m is the length of the right leg time series. In order to align them in the time dimension, it is necessary to construct an n×m matrix grid, and the position (i, j) of the matrix is ​​used to store point a i and point b j The Euclidean distance, d(a i ,b j) is smaller, the higher the similarity. In the n×m matrix grid, the continuous set of elements constitutes a path, which is defined as a regular path and represented by W:

[0059] W=(ω 1 ,ω 2 ,...,ω k )(max(n,m)≤k≤m+n-1) (1)

[0060] Among them, ω k Represents the kth element of W.

[0061] Each continuous point can form a curve, so to find the path with the shortest cumulative distance:

[0062]

[0063] K represents the number of path points, which is used to compensate for regular paths W of different lengths.

[0064] Initialize the distance matrix D and set D(1,1) to the distance between the first element of the left leg time series A and the right leg time series B. Starting from the starting point (1,1), the minimum cumulative value D(n,m) is calculated iteratively, which represents the shortest cumulative distance DTW(A,B) between the left leg time series A and the right leg time series B.

[0065] 1.2 Gait consistency

[0066] Gait consistency refers to the degree of change in gait symmetry when the robot repeatedly performs the same movement under the same conditions. The smaller the degree of change, the better the gait consistency. The present invention calculates the variance of the gait symmetry obtained by repeated movements to obtain the gait consistency. The calculation formula is as follows:

[0067]

[0068] Where n represents the number of times the movement is repeated. is the average value of gait symmetry.

[0069] 1.3 Real-time action recognition

[0070] The present invention proposes to combine a large public data set in the field of human motion recognition with a small amount of humanoid robot gait data to train a gait recognition model, and deploy the pre-trained model on the humanoid robot to recognize the gait of the humanoid robot in real time, thereby intuitively presenting the robot state. The pre-trained model can recognize actions including standing, sitting, lying, walking, running, going up and down slopes, and going up and down stairs.

[0071] The present invention adopts the ConvLSTM model to classify actions. The model is mainly composed of a convolutional layer, an LSTM layer, a MaxPool2D layer and a fully connected layer, wherein the output layer is composed of 9 neurons for classifying and representing nine actions. Specifically, the proposed Convolutional LSTM model includes 2 Convolutional 2D layers, a maximum pooling layer, 4 bidirectional LSTM layers and an output layer. Among them, the Conv1D layer uses the relu activation function, the kernel size is 5, the number of convolution kernels is set to 128, and the fully connected layer uses the softmax activation function. Figure 2 The ConvLSTM model architecture proposed in the present invention is demonstrated.

[0072] 2D-CNN uses a two-dimensional convolution kernel to perform convolution with gait data, and the process is shown in formula (3-4).

[0073]

[0074] Among them, P i and Q i are the length and width of the convolution kernel, respectively. is the value of the convolution kernel at the point (p,q) connected to the mth feature map, is the convolution value of the jth feature map at position (x, y) of the i-th layer network, b ij is the bias term, and tanh() represents the hyperbolic tangent function.

[0075] LSTM has shown strong performance in processing time series data tasks. The following is a brief introduction to the LSTM network. LSTM has three gate structures: forget gate, t , input gate i t And the output gate o t The forget gate is used to filter the previous hidden state h t-1 and input x t The content that needs to be retained, when the input gate reads into the memory element C t The output gate selects which information is passed to the next unit.

[0076] f t =σ(W f [h t-1 ,x t ]+b f ) (5)

[0077] i t =σ(W t [h t-1 ,x t ]+b i ) (6)

[0078] o t =σ(W o [h t-1 ,x t ]+b o ) (7)

[0079]

[0080]

[0081] h t =o t ⊙tanh(C t ) (10)

[0082] Among them, W represents the weight parameter, b is the bias term, σ is the sigmoid activation function, represents the updated value of the unit state, ⊙ is the Hadamard product operator, h t is the hidden state at time t.

[0083] The present invention uses a sliding window mechanism to detect robot movements in real time. Figure 3 As shown in the figure, the window length is set to 2s, the window moves every 0.5s, and the gait signal is intercepted with 25% overlap. The intercepted motion data is input into the ConvLSTM model to recognize actions in real time.

[0084] 1.4 Health Assessment

[0085] Nervous system diseases can seriously affect human gait. The present invention collects the gait of healthy subjects and patients to build a human gait data set for training a health assessment model. After training, this model is used to measure the gait health of a humanoid robot, so that the robot gait can be adjusted in time. The health assessment model has the same structure as the motion recognition model, except that the number of output neurons is set to 1.

[0086] 1.5 Gait phase assessment

[0087] Gait phase recognition requires higher real-time performance, so the present invention uses a lightweight network model to achieve this task. The present invention divides gait phases into 7 categories, including 6 categories in a complete gait cycle: heel strike, foot flat, mid-stance, heel off, toe off, swing cycle, and other categories (non-walking movements).

[0088] Step 2: Reinforcement Learning Strategy Setting

[0089] 2.1 State Space

[0090] The state space is a comprehensive description of the robot at the current moment, covering the robot's own perception state. The state space selected by the present invention includes

[0091]

[0092] Among them, q j and denote the position and velocity of the jth joint of the humanoid robot, p i ,q i ,v i ,w i They respectively represent the relative position, direction, linear velocity and angular velocity of the i-th joint of the humanoid robot relative to the root node (pelvis) in the local coordinate system, h represents the height of the pelvis of the humanoid robot from the ground, sy, co, ac, ph, and he represent the gait symmetry, gait consistency, action, gait phase, and gait health of the robot, respectively.

[0093] 2.2 Action Space

[0094] The present invention selects all the degrees of freedom of the robot as the action space, and by precisely regulating each degree of freedom of a joint, the robot can be controlled to achieve a desired posture or position, thereby executing complex tasks or action sequences.

[0095] 2.3 Network Structure

[0096] The present invention uses the Proximal Policy Optimization (PPO) algorithm as the core framework of reinforcement learning. The framework adopts an actor-critic structure, which includes a policy network and a value network. The policy network is used to generate and execute corresponding behavioral actions, while the value network is responsible for evaluating the value brought by these actions. In the present invention, both the policy network and the value network are constructed using a fully connected network.

[0097] Step 3: Pre-training model deployment. After freezing the trained algorithm model parameters, it is used in the task training of the humanoid robot. The pre-training model is used to directly infer the robot's gait symmetry, gait consistency, actions, gait phase, and health, so that the robot can learn higher-level motion information and help the robot perform better actions. Among them, the estimated gait symmetry, consistency, and health status can be used as important parameters to help the robot adjust to a reasonable gait, and the estimated robot action can provide prior information for the next step of the robot's control.

[0098] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for analyzing gait of a humanoid robot, characterized in that the steps include: The collected gait data is used to pre-train a gait analysis and recognition model to evaluate the gait parameters of the humanoid robot, wherein the gait parameters include: gait symmetry, gait consistency, movement, gait phase, and gait health; the gait consistency refers to the degree of change in gait symmetry when the robot repeatedly performs the same movement under the same conditions; The gait analysis and recognition model calculates the waveform similarity of the left and right leg time domain signals through a dynamic time warping algorithm to obtain gait symmetry, as follows: Based on the left leg time series A={a1,a2,...,a i ,...,a n } and the right leg time series B={b1,b2,...,b j ,...,b m }, construct an n×m matrix grid, the position (i, j) of the matrix is ​​used to store point a i and point b j The Euclidean distance of In the matrix grid, the continuous set of elements constitutes the path as a regular path W; Find the path with the shortest cumulative distance: Among them, K represents the number of path points, ω k represents the kth element of the regularized path W; Initialize the distance matrix D and set D(1,1) to the distance between the first element of the left leg time series A and the right leg time series B; Starting from the starting point, the minimum cumulative value D(n,m) obtained by iterative calculation is the shortest cumulative distance DTW(A,B) between the left leg time series and the right leg time series; Gait parameters are estimated in real time based on the pre-trained gait analysis and recognition model.

2. A humanoid robot gait analysis method according to claim 1, characterized in that: The gait analysis and recognition model evaluates the consistency of gait by calculating the variance of gait symmetry obtained from multiple repeated movements.

3. A humanoid robot gait analysis method according to claim 1, characterized in that: The real-time action recognition of the gait analysis and recognition model adopts a ConvLSTM-based action classification model to classify actions. The action classification model consists of a convolutional layer, an LSTM layer, a MaxPool2D layer, and a fully connected layer. The output layer of the action classification model is provided with a plurality of neurons for classifying and representing different types of actions.

4. A humanoid robot gait analysis method according to claim 3, characterized in that: The real-time action recognition uses a sliding window mechanism to detect the robot's actions in real time, intercepts gait signals based on set window length, window movement frequency and overlap rate, and inputs the intercepted motion data into the ConvLSTM-based action classification model to recognize actions in real time.

5. The method for analyzing gait of a humanoid robot according to claim 1, characterized in that: The gait health assessment of the gait analysis and recognition model is as follows: The gaits of healthy subjects and patients are collected to construct a human gait dataset and train a health assessment model; the health assessment model adopts a ConvLSTM model, which is composed of a convolutional layer, an LSTM layer, a MaxPool2D layer and a fully connected layer, wherein the number of neurons in the output layer of the health assessment model is set to 1.

6. A humanoid robot gait analysis method according to claim 1, characterized in that: The classification of the gait phases includes: heel strike, foot flat, mid-stance, heel lift, toe lift, swing cycle and non-walking action in a complete gait cycle; the gait phases are evaluated using a fully connected network model.

7. A robot control method, characterized in that: The method comprises the following steps: Establishing a pre-trained gait analysis and recognition model based on the humanoid robot gait analysis method as described in any one of claims 1 to 6; deploying the pre-trained gait analysis and recognition model into a humanoid robot control strategy; The gait parameters estimated in real time by the gait analysis and recognition model are added to the state space of the humanoid robot control strategy, and the control strategy generates and executes corresponding behavioral actions.

8. A robot control method according to claim 7, characterized in that: The state space includes: the position and velocity of the humanoid robot's joints, the relative position, direction, linear velocity and angular velocity of the humanoid robot's joints relative to the root node in the local coordinate system, the height of the humanoid robot's pelvis from the ground, and the robot's gait symmetry, gait consistency, movement, gait phase and gait health.

9. A robot control method according to claim 7, characterized in that: The humanoid robot control strategy uses the proximal policy optimization algorithm as the core framework of reinforcement learning and adopts a fully connected network to construct a policy network and a value network; The strategy network is used to generate and execute corresponding behavioral actions, and the value network is responsible for evaluating the value brought by the actions generated by the strategy network.

Images