Robot environment contact action element network modeling method

Abstract

The invention belongs to the field of robot force perception, and particularly relates to an element network modeling method for robot environment contact action modeling. The element network uses the speed of the robot as input and the contact force as output, and abundant contact information of the robot in the movement process is fully utilized. Through iteration of element network parameter training and structure training, the predictive force gradually fits the contact force, and the network structure is converged to the simplest. The final element network is a contact model between the robot and the environment, and has the advantages of high precision and high interpretability. According to the method, the defect that precision and interpretability cannot be considered at the same time in contact action modeling is overcome, and a new solution is provided for force perception of the robot.

Description

Technical Field

[0001] The present invention belongs to the technical field of robot force perception, and in particular relates to a component network modeling method for robot environment contact action modeling. Background Art

[0002] With the development of robotics technology, the application scenarios of robots are becoming more and more complex. In tasks such as disaster relief, pipeline maintenance, and wilderness exploration, legless robots with unique mobility and flexibility have become the best choice for performing tasks. Thanks to the diverse movement modes of legless robots, they can maintain efficient movement on different terrains. Taking post-disaster rescue as an example, the creeping mode ensures that the legless robot can shuttle through ruins and rubble, and the snake-like crawling can maximize its movement speed on open and stable ground. Therefore, how to identify the terrain is crucial to fully utilize the movement capabilities of legless robots.

[0003] By integrating various sensors, robots can perceive their direct environment through visual, auditory or tactile feedback. Currently, computer vision is the most rapidly developing and most common method for environmental perception. However, the visual system mainly extracts geometric information about the environment, including texture, depth or color features, which may lead to cognitive biases in environments such as poorly lit underground pipes or ruins full of visual obstacles. In contrast, there is no similar problem in using touch to recognize the environment.

[0004] After obtaining force feedback, how to process high-dimensional time-varying contact signals is a key step in enabling the robot to understand, analyze and predict the contact environment. Due to the complexity of the contact behavior between the robot and the environment, it is difficult to accurately describe the mechanical characteristics through first principles, so data-driven modeling through machine learning methods has become the preferred solution. However, most data-driven modeling methods rely on "black box" models with hidden internal logic. Such models only provide input-output mapping, lack interpretability, cannot judge the essential characteristics of the contact environment, and are prone to failure in unfamiliar environments. Therefore, it is necessary to propose a physically interpretable environmental contact action modeling method for the robot's force perception. Summary of the invention

[0005] The object of the present invention is to provide a component network modeling method for modeling robot-environment contact interaction. The present invention proposes a component network for establishing an interpretable contact model between the robot and the environment.

[0006] The present invention proposes a component network modeling method for robot-environment contact modeling, which uses the robot's motion information as input and the predicted force of the environmental contact as output. By training the component network, its predicted force fits the real contact force. When the fitting accuracy reaches a certain value, the model of the component network can be regarded as the contact model between the robot and the environment; the specific steps are as follows: (1) Initializing the structure of the component network, determining that the input in the component network is the movement speed of the robot, and the output is the contact force generated during the movement; the component network is formed by connecting the required components in parallel to form a component layer, connecting several component layers in series to form a component group, and then connecting each component group in parallel to form a network; (2) Determine the parameter boundaries of the component network, randomly initialize the values ​​of each component, and use a heuristic optimization method based on differentiation to train the component parameters; (3) The network structure is trained according to the optimal component parameters. At the component group, component layer, and component levels, units whose contribution to the component network prediction power is less than the threshold are deleted. Then, homogeneous components are fused to ensure the simplicity of the component network structure. Some components are reactivated through a random wake-up mechanism to avoid accidental deletion of important components. (4) Repeat steps (2) and (3) until the component network structure converges, and derive the optimal parameters of the network structure and each component in the network as the contact model of the robot moving in the environment.

[0007] In the present invention, the required element in step (1) is any one or more of a spring element, a damping element or a slider element.

[0008] In the present invention, the optimal parameters of each element in step (4) are specifically any one or more of the stiffness parameter of the spring element, the viscous friction coefficient of the damping element, or the dry friction coefficient of the slider element.

[0009] The beneficial effects of the present invention are: 1. Compared with first-principles modeling, component networks have higher accuracy and can describe complex nonlinear phenomena in contact.

[0010] 2. By introducing interpretability, the contact model can characterize the characteristics of the environment, thereby guiding the design of the robot's structure or motion mode.

[0011] 3. The commonalities or differences between different contact environments can be reflected through the component network model, so that in unfamiliar environments, the robot can still switch to the appropriate mode according to the contact model.

[0012] 4. The method of the present invention only requires a force sensor, and the motion information can be obtained through the robot's host computer or an additional displacement sensor. Few sensors are required and the implementation is simple. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 shows the general architecture of a component network.

[0014] Figure 2 is a flowchart of component network training. DETAILED DESCRIPTION

[0015] Example 1: Modeling the contact effects of segments of a worm-like legless mobile robot in eight environments, including grass, sand, and carpet. In the contact force acquisition experiment, four groups of steel balls of different numbers were applied to the segments to simulate different masses, and they moved in seven different motion modes to ensure the versatility of the contact model. The robot segments obtain contact information through six-dimensional force sensors during movement, and record the current speed and displacement data in real time through displacement sensors. The experimental results show that the contact model established using the component network can fit the contact force exerted on the robot, and the mean square error in each environment is less than 0.1. When using the component network as an environmental feature for environmental perception, the recognition accuracy rate reached 87%, proving the dual advantages of the component network in accuracy and interpretability.

[0016] Figure 1 It is a general architecture of component networks. The mechanical component network consists of three types of components: sliders, dampers, and linear springs, which respectively represent the common dry friction, viscous friction, and elastic deformation in contact. These three types of components are first connected in parallel to form modules; each module is connected in series to form a group; and the group is then connected in parallel to form a network.

[0017] Figure 2 The training process of the component network is divided into two parts: component parameter training and network structure training. In the component parameter training, a heuristic optimization method based on differential evolution is designed to optimize parameters through crossover mutation between populations and individuals, and to adaptively adjust hyperparameters. Specifically, the individuals and hyperparameters in the population are first randomly initialized, and then the individual parameters and population are updated according to the initialization results. After each population update, the population size is used to determine whether the optimization termination condition is met. If not, crossover and mutation operations are performed to evolve the population. Individuals with smaller losses will be added to the new population and update parameters, and individuals with larger losses will be placed in external archives as references for mutation; if the optimization termination condition is met, network structure training is performed. According to the results of parameter training, we initialize the network structure, and then calculate the contribution of each component group, component layer, and component to the network prediction power, delete units with small contributions, and merge homogeneous components in series or parallel. If the network structure converges to the optimal at this time, the component network is directly output, otherwise some component modules are randomly awakened and the next round of iteration is performed.

Claims

1. A component network modeling method for modeling robot-environment contact effects, characterized in that: Using the robot's motion information as input and the predicted force of the contact with the environment as output, the component network is trained to make its predicted force fit the real contact force. When the fitting accuracy reaches a certain value, the model of the component network can be regarded as the contact model between the robot and the environment; the specific steps are as follows: (1) Initializing the structure of the component network, determining that the input in the component network is the movement speed of the robot, and the output is the contact force generated during the movement; the component network is formed by connecting the required components in parallel to form a component layer, connecting several component layers in series to form a component group, and then connecting each component group in parallel to form a network; (2) Determine the parameter boundaries of the component network, randomly initialize the values ​​of each component, and use a heuristic optimization method based on differentiation to train the component parameters; (3) The network structure is trained according to the optimal component parameters. At the component group, component layer, and component levels, units whose contribution to the component network prediction power is less than the threshold are deleted. Then, homogeneous components are fused to ensure the simplicity of the component network structure. Some components are reactivated through a random wake-up mechanism to avoid accidental deletion of important components. (4) Repeat steps (2) and (3) until the component network structure converges, and derive the optimal parameters of the network structure and each component in the network as the contact model of the robot moving in the environment.

2. The component network modeling method for robot environment contact modeling according to claim 1 is characterized by: The required element in step (1) is any one or more of a spring element, a damping element or a slider element.

3. The component network modeling method for robot environment contact action modeling according to claim 1 is characterized by: The optimal parameters of each element in step (4) are specifically any one or more of the stiffness parameter of the spring element, the viscous friction coefficient of the damping element, or the dry friction coefficient of the slider element.

4. The component network modeling method for robot environment contact modeling according to claim 1 is characterized by: In step (2), the heuristic optimization method based on differential evolution optimizes parameters through crossover mutation between population and individuals, and adaptively adjusts hyperparameters. Specifically, individuals and hyperparameters in the population are first randomly initialized, and then individual parameters and population are updated according to the initialization results. After each population update, it is determined whether the optimization termination condition is met based on the population size. If not, crossover and mutation operations are performed to evolve the population. Individuals with smaller losses are added to the new population and their parameters are updated, while individuals with larger losses are placed in external archives as references for mutation. If the optimization termination condition is met, network structure training is performed. According to the results of parameter training, we initialize the network structure, then calculate the contribution of each component group, component layer and component to the network prediction power, delete the units with small contribution, and merge the homogeneous components in series or parallel; If the network structure converges to the optimal state at this time, the component network is directly output; otherwise, some component modules are randomly awakened and the next round of iteration is performed.

Images