CN117444978B - Position control method, system and equipment for pneumatic soft robot - Google Patents

Abstract

The present invention discloses a position control method for a pneumatic soft robot and its system and equipment. The method obtains an initial data set including several states and corresponding action data groups through a random strategy, and then uses the above data set as a training input, uses a differential variable representing the difference between the Gaussian process function mapping and the actual state as a training target, and uses the posterior predictive distribution of the differential variable obtained by Bayesian inference to represent the transfer dynamics function in the dynamics model. Finally, the learned dynamics model is used to calculate the predicted state distribution, and the strategy parameters are optimized with the goal of minimizing the expected long-term loss to achieve an update strategy. The present invention uses a probability model to replace the actual dynamics model, thereby reducing the model deviation and solving the problem of the existing method over-relying on the precise dynamics model.

Description

A method and system for controlling the position of a pneumatic soft robot

Technical Field

The present invention belongs to the technical field of intelligent robots, and in particular relates to a position control method of a pneumatic soft robot and a system and equipment thereof.

Background technique

Pneumatic soft robots can perform actions that existing robots cannot, especially in some dangerous confined spaces. Due to the complex structure of pneumatic soft robots, the problems of modeling and position control have not yet been solved and need to be further explored and developed.

Position control schemes are basically divided into two categories, static controllers and dynamic controllers. Static controllers are always built on the kinematic models of these robots, which are based on steady-state assumptions, which leads to limitations in their speed, efficiency, and reachability. Dynamic controllers are designed to exploit and take advantage of the complex dynamic characteristics of these systems. However, this is not simple, as building an accurate dynamic model is much more complicated than building a kinematic model.

Summary of the invention

The purpose of the present invention is to provide a position control method and system and equipment for a pneumatic soft robot, which utilizes a probabilistic model to replace the actual dynamic model, thereby reducing model deviation and solving the problem of over-reliance on precise dynamic models in existing methods.

The objective of the present invention is achieved through the following technical solutions:

A method for controlling the position of a pneumatic soft robot comprises the following steps:

Step S1: Obtaining the initial experience pool:

Applying a random strategy to the controller, using sensors in the system to obtain an initial data set containing several robot states and corresponding controller action data sets;

Step S2: Dynamic model learning:

The initial data set obtained in step S1 is used as training input, and the difference variable representing the difference between the Gaussian process function map and the actual state is used as the training target; for a known strategy, the posterior predictive distribution of the difference variable obtained by Bayesian inference is used to represent the transfer dynamics function in the dynamics model;

Step S3: Policy update:

The dynamic model learned in step S2 is used to calculate the predicted state distribution, evaluate the long-term loss, and optimize the strategy parameters with the goal of minimizing the expected long-term loss to achieve the updated strategy.

A position control system for a pneumatic soft robot comprises a controlled pneumatic soft robot unit, an input and output signal acquisition unit, and a data processing and transmission unit, wherein:

The pneumatic soft robot unit includes at least one joint of a control object for driving the robot to move;

The input and output signal acquisition unit can at least be used to collect a data set of the position control method of the pneumatic soft robot;

The data processing and transmission unit can at least be used to run and store the position control method and signal of the pneumatic soft robot.

A position control device for a pneumatic soft robot comprises at least one processor and a memory in communication with the at least one processor, wherein:

The memory stores instructions executable by the at least one processor;

The instructions are executed by the at least one processor so that the at least one processor can execute the position control method of the pneumatic soft robot.

In the present invention, the position control device of the above-mentioned pneumatic soft robot also includes a computer-readable storage medium, wherein: the computer-readable storage medium stores computer-executable instructions; the computer-executable instructions are used to enable a computer to execute the position control method of the above-mentioned pneumatic soft robot.

Compared with the prior art, the present invention has the following advantages:

1. The position control method of the pneumatic soft robot provided by the present invention uses a probabilistic model to replace the actual dynamic model, thereby reducing the model deviation. At the same time, the model obtained based on probabilistic reasoning can express and represent the uncertainty of the dynamic model, and incorporate the uncertainty of the model into long-term planning and decision-making.

2. The position control method of the pneumatic soft robot provided by the present invention utilizes the characteristics that the prediction result of the measured input is seriously dependent on the training samples near it, and selects training samples in the area near the test input to complete local GP regression, which not only ensures the prediction accuracy but also reduces the training time of the model.

3. The position control method of the pneumatic soft robot provided by the present invention combines physical interaction with simulation learning. The state transition model is established using the interaction data, and the simulation strategy is optimized using this model to reduce the risk and cost of physical interaction.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of position control of a pneumatic soft robot.

FIG. 2 is a schematic diagram of the structure of the position control system of the pneumatic soft robot.

Detailed ways

The technical solution of the present invention is further described below in conjunction with the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

The present invention provides a position control method for a pneumatic soft robot. An initial data set containing a number of states and corresponding action data groups is obtained through a random strategy. The data set is then used as a training input. A differential variable representing the difference between a Gaussian process function map and an actual state is used as a training target. The posterior predictive distribution of the differential variable obtained by Bayesian inference is used to represent the transfer dynamics function in the dynamics model. Finally, the learned dynamics model is used to calculate the predicted state distribution, and the strategy parameters are optimized with the goal of minimizing the expected long-term loss to achieve an update strategy. As shown in FIG1 , the method comprises the following steps:

Step S1: Obtaining the initial experience pool:

Apply a random strategy to the controller and use the sensors in the system to obtain an initial data set containing several robot states and corresponding controller action data sets. The specific steps are as follows:

Apply a random strategy to the control object in the robot system and change the controller's actions at the same time. And the corresponding robot state/> Record as a set of data, repeatedly apply random strategies to obtain a data set containing several state-action data sets/> Where n represents the number of times the random strategy is applied.

Step S2: Dynamic model learning:

The initial data set obtained in step S1 is used as training input, and the differential variable representing the difference between the Gaussian process function map and the actual state is used as the training target; for known strategies, the posterior predictive distribution of the differential variable obtained by Bayesian inference is used to represent the transfer dynamics function in the dynamic model.

Specific steps are as follows:

Step S21: Use the data ( _xi , _ui ), i=1, ..., n, obtained in step S1 as training input for a dynamics model established based on a Gaussian process.

In order to reduce the training time of the dynamics model, the training input needs to be preprocessed. The specific steps are as follows:

(1) Obtain the distance between the target point and the training data set;

(2) Calculate the distance and sort it;

(3) Select K points closest to the training data set, where: there is no specific requirement for the value of K. If the K value is small, the approximation error is low, but the estimation error is high; on the contrary, if the K value is large, the approximation error is high, but the estimation error is low. Therefore, in actual use, it is necessary to make continuous adjustments according to the control effect to determine the final applicable value;

(4) The selected K points are used as the new training data set and local Gaussian process regression is performed.

Step S22: For the M-dimensional state space, the difference variable between the function mapping f _m ( _xi , _ui ) of the m-th Gaussian process and the true state _xim is used as the training target:

Δx _im = f _m ( x _i , u _i ) - x _im , m = 1, ..., M

Among them, Δx _im is also called the training target of the Gaussian process.

Step S23: for a given deterministic input (x _* , u _* ), the mean and variance of the Gaussian subsequent state distribution p( _fm (x _* , u _* )) are:

Wherein, the definitions of x _* and u _* are the same as those in S1, and represent the state of the robot and the action of the controller, respectively. For the position control method of the pneumatic soft robot of the present invention, the controller action here can be the driving signal generated by each pneumatic driving element, such as the pressure of a proportional valve, and the robot state here is the (x, y, z) position information of the robot end effector; and var are the signs of the mean and variance respectively;

The prediction distribution composed of the above two equations describes the uncertainty relationship of the implicit function.

Step S24: Integrate the M Gaussian processes and use Bayesian inference to obtain the transfer kinetics function in the kinetic model represented by the posterior predictive distribution of the differential variable as follows:

in, is a general expression of Gaussian distribution, where μ is the mean of the distribution and σ ² is the variance of the distribution. In this formula, the value of S23 is used to represent the distribution. Replace with var formula.

Step S3: Policy update:

The dynamic model learned in step S2 is used to calculate the predicted state distribution, evaluate the long-term loss, and optimize the strategy parameters with the goal of minimizing the expected long-term loss to achieve the updated strategy. The specific steps are as follows:

Step S31: In the actual system, the control signal (pressure or flow) has certain constraints, that is, u∈[-u _max ,u _max ]. Therefore, a sine function is used to adjust the initial strategy. Constrained, and then limit the amplitude of strategy π, the final strategy π can be described as/> Among them, u _max represents the maximum control signal of the actual controller, such as the maximum pressure that a proportional valve can reach.

Step S32: Calculate the predicted state distribution p(x ₁ ), ..., p(x _T ) based on the dynamic model obtained in step S2 to assess the long-term loss. The deterministic conjugate gradient minimization method is used to update the policy parameters, and the following saturation cost function is used as the penalty term:

Among them, the subscript target represents the target position;

Step S33: Use the gradient strategy search method to search for the optimal strategy π ^* :

FIG2 shows an embodiment of a position control system of a pneumatic soft robot, wherein the position control system 2 of the pneumatic soft robot includes at least one joint, such as a first joint 21 of the robot, a second joint 22 of the robot, a robot end effector 23, a robot end position collector 24, and a data processing and transmission unit 25, wherein:

The first joint 21 and the second joint 22 of the robot contain at least one control object, such as the first joint pneumatic element (211a, 211b, 211c) of the robot, the second joint pneumatic element (221a, 221b, 221c) of the robot, and a driving element matched with the control object, such as the first joint pneumatic driving element (212a, 212b, 212c) of the robot, and the second joint pneumatic driving element (222a, 222b, 222c) of the robot;

The robot end position collector 24 is used to collect the position information output by the robot end effector 23 in real time and upload it to the data processing and transmission unit 25;

The robot first joint pneumatic drive element (212a, 212b, 212c) and the robot second joint pneumatic drive element (222a, 222b, 222c) should include a signal acquisition module for detecting the control signal input to the robot system and uploading it to the data processing and transmission unit 25;

The data processing and transmission unit 25 can at least be used to run and store the control method and signals in the above steps S1 to S3.

At the same time, the above-mentioned data processing and transmission unit 25 can be used as a position control device of a pneumatic soft robot, which includes at least one processor and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the position control method of the pneumatic soft robot as described in the above-mentioned steps S1 to S3.

Claims (4)

1. A method for controlling the position of a pneumatic soft robot, characterized in that the method comprises the following steps: Step S1: Obtaining the initial experience pool: Apply a random strategy to the controller and use the sensors in the system to obtain an initial data set containing several robot states and corresponding controller action data sets. The specific steps are as follows: Apply a random strategy to the control object in the robot system, record the controller's action u and the corresponding robot state x at the same time as a set of data, and repeatedly apply the random strategy to obtain a data set containing several state-action data sets. Step S2: Dynamic model learning: The initial data set obtained in step S1 is used as the training input, and the differential variable representing the difference between the Gaussian process function map and the actual state is used as the training target; for a known strategy, the posterior predictive distribution of the differential variable obtained by Bayesian inference is used to represent the transfer kinetic function in the kinetic model; the specific steps are as follows: Step S21: Use the data ( _xi , _ui ) obtained in step S1 as the training input of the dynamic model established based on the Gaussian process, i = 1, ..., n, u is the action of the controller, x is the robot state, and n represents the number of times the random strategy is applied; in order to reduce the training time of the dynamic model, the training input needs to be preprocessed, and the specific steps are as follows: (1) Obtain the distance between the target point and the training data set; (2) Calculate the distance and sort it; (3) Select the K points closest to the training data set; (4) The selected K points are used as new training data sets and local Gaussian process regression is performed; Step S22: For the M-dimensional state space, the difference variable between the function mapping f _m ( _xi , _ui ) of the m-th Gaussian process and the true state _xim is used as the training target: Δx _im = f _m ( x _i , u _i ) - x _im , m = 1, ..., M Among them, Δx _im is also called the training target of the Gaussian process; Step S23: for a given deterministic input (x _* , u _* ), the mean and variance of the Gaussian subsequent state distribution p( _fm (x _* , u _* )) are: in, and var are the signs of the mean and variance respectively; Step S24: Integrate the M Gaussian processes and use Bayesian inference to obtain the transfer kinetics function in the kinetic model represented by the posterior predictive distribution of the differential variable as follows: Step S3: Policy update: The dynamic model learned in step S2 is used to calculate the predicted state distribution, evaluate the long-term loss, and optimize the strategy parameters with the goal of minimizing the expected long-term loss to implement the update strategy. The specific steps are as follows: Step S31: Use a sine function to calculate the initial strategy To constrain the strategy π, the strategy π is finally described as Among them, u _max represents the maximum control signal of the actual controller; Step S32: Calculate the predicted state distribution p(x ₁ ), ..., p(x _T ) based on the dynamic model obtained in step S2 to assess the long-term loss. The deterministic conjugate gradient minimization method is used to update the policy parameters, and the following saturation cost function is used as the penalty term: Among them, the subscript target represents the target position; Step S33: Search for the optimal strategy π ^* using a gradient strategy search method: 2. A position control system for a pneumatic soft robot, characterized in that the system comprises a controlled pneumatic soft robot unit, an input and output signal acquisition unit, and a data processing and transmission unit, wherein: The pneumatic soft robot unit includes at least one joint of a control object for driving the robot to move; The input and output signal acquisition unit is at least used to collect a data set of the position control method of the pneumatic soft robot; The data processing and transmission unit is at least used to run and store the position control method and signal of the pneumatic soft robot according to claim 1. 3. A position control device for a pneumatic soft robot, characterized in that the device comprises at least one processor and a memory in communication with the at least one processor, wherein: The memory stores instructions executable by the at least one processor; The instructions are executed by the at least one processor so that the at least one processor can execute the position control method of the pneumatic soft robot according to claim 1. 4. The position control device of a pneumatic soft robot according to claim 3 is characterized in that the device also includes a computer-readable storage medium, wherein: the computer-readable storage medium stores computer-executable instructions; the computer-executable instructions are used to enable a computer to execute the position control method of the pneumatic soft robot described in 1.