A robot gait optimization method

Abstract

In order to improve the stability control effect of a bipedal robot in a collaborative state, the present invention provides a robot gait optimization method for optimizing the gait of a first bipedal robot, including: establishing a neural network model based on a neural cell layer, training the gait data of the first bipedal robot obtained by a gait measurement device, and training the gait balance of the first bipedal robot; optimizing the gait of the first bipedal robot according to the balance result and the gait data of the second bipedal robot. The bipedal robot of the present invention can not only improve the training depth of the gait control model based on its own or other bipedal robots' experience data, but also can effectively cooperate with other bipedal robots to realize mutual experience sharing, thereby achieving the effects of jointly carrying the same goods, quickly adapting to new environments to provide more important tasks with sufficient runtime, etc.

Description

Technical Field

[0001] The present invention relates to the field of automatic control technology, and more specifically, to a robot gait optimization method. Background Art

[0002] With the advancement of science and technology, bipedal robots have developed rapidly and are widely used in many fields. At present, most buildings and tools are designed according to human height and body shape, so bipedal robots have better flexibility as a robot platform. At the same time, the stability control of the gait of bipedal robots is the premise and basis for the smooth walking of robots. Gait refers to the relationship between various joints in time and space during standing or walking, which can be described by the movement trajectory of the joints.

[0003] Existing studies on the stability of bipedal robot gait are all based on the zero moment point (ZMP) method, which establishes a mathematical model of the bipedal robot and derives the control law based on the fact that the ZMP must fall in the stable area, such as the range of the robot's foot. However, factors such as road conditions may cause the robot to fail to walk or even fall. This phenomenon is a fatal disaster when multiple bipedal robots collaborate. Summary of the invention

[0004] In order to improve the stability control effect of a bipedal robot in a collaborative state, the present invention provides a robot gait optimization method for optimizing the gait of a first bipedal robot, comprising: establishing a neural network model based on a neural cell layer, training the gait data of the first bipedal robot obtained using a gait measurement device, and training the gait balance of the first bipedal robot; optimizing the gait of the first bipedal robot according to the balance result and the gait data of a second bipedal robot.

[0005] Furthermore, the robot's feet, ankle joints and knee joints all have three degrees of freedom.

[0006] Furthermore, the gait data includes angular velocity and acceleration, and the angular velocity and acceleration are the angular velocity and acceleration in the forward direction.

[0007] Furthermore, the gait measurement device includes a gyroscope and an accelerometer.

[0008] Furthermore, the training of gait data includes:

[0009] Obtaining road surface information data, wherein the road surface information data at least includes route information and experience data in an electronic map that is connected to the current location;

[0010] Gait data is obtained, and the pre-built neural network model is trained based on the gait data to obtain a first-level training model:

[0011]

[0012]

[0013] Where i represents the degree of freedom, which can be 1, 2 or 3; p and q represent a positive integer randomly selected between 1 and 10, and when the subscripts of the above two equations have no corresponding physical meaning, they are taken as 0;

[0014] The first-level training model is used to perform deep learning on the empirical data, and the learning results are used to train the Hopfield network to obtain the second-level training model.

[0015] Further, training gait balance includes:

[0016] Obtaining gait data of at least one second bipedal robot, wherein the number of the second bipedal robots is greater than or equal to 3;

[0017] The secondary training model of the first bipedal robot is used to perform deep learning on the gait data of each second bipedal robot, and the learning results are summarized. The final learning results are used for sparse autoencoder training to obtain the third-level training model of the first bipedal robot.

[0018] Furthermore, optimizing the gait of the first bipedal robot according to the balancing result and the gait data of the second bipedal robot includes: obtaining a second bipedal robot that is farthest away from the first bipedal robot in the walking direction, and performing deep learning on the secondary training model of the first bipedal robot based on the gait data of the second bipedal robot, and the learning result is used for sparse autoencoder training to obtain a four-level training model of the first bipedal robot; for the action setting of each degree of freedom of the bipedal robot, ankle joints and knee joints of the first bipedal robot, the action parameters are output by the model obtained after training the four-level training model with the gait data of the first bipedal robot.

[0019] Furthermore, the first-level training model is selected by using a random selection method in a Monte Carlo search tree optimization algorithm.

[0020] Furthermore, the experience data is gait data from when the first bipedal robot or the second bipedal robot has previously passed a route that is most similar to the route information.

[0021] The beneficial effects of the present invention are that the bipedal robot can not only improve the training depth of the gait control model based on its own or other bipedal robots' experience data, but also can effectively cooperate with other bipedal robots to share experience with each other, thereby achieving the effects of jointly carrying the same cargo, quickly adapting to new environments to provide sufficient runtime for more important tasks, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A flowchart of the method is shown. DETAILED DESCRIPTION

[0023] A robot gait optimization method is used to optimize the gait of a first biped robot, comprising: establishing a neural network model based on a neural cell layer, training the gait data of the first biped robot obtained using a gait measurement device, and training the gait balance of the first biped robot; and optimizing the gait of the first biped robot according to the balance result and the gait data of a second biped robot.

[0024] Preferably, the feet, ankle joints and knee joints of the robot all have three degrees of freedom.

[0025] Preferably, the gait data includes angular velocity and acceleration, and the angular velocity and acceleration are the angular velocity and acceleration in the forward direction.

[0026] Preferably, the gait measurement device comprises a gyroscope and an accelerometer.

[0027] Preferably, the training of gait data includes:

[0028] Obtaining road surface information data, wherein the road surface information data at least includes route information and experience data in an electronic map that is connected to the current location;

[0029] Gait data is obtained, and the pre-built neural network model is trained based on the gait data to obtain a first-level training model:

[0030]

[0031]

[0032] Where p and q represent a positive integer randomly selected between 1 and 10, and when the subscripts in the above two equations have no corresponding physical meaning, they are taken as 0;

[0033] The first-level training model is used to perform deep learning on the empirical data, and the learning results are used to train the Hopfield network to obtain the second-level training model.

[0034] Preferably, training gait balance includes:

[0035] Obtain gait data of at least one second bipedal robot, the number of the second bipedal robots being greater than or equal to 3:

[0036] The secondary training model of the first bipedal robot is used to perform deep learning on the gait data of each second bipedal robot, and the learning results are summarized. The final learning results are used for sparse autoencoder training to obtain the third-level training model of the first bipedal robot.

[0037] Preferably, optimizing the gait of the first bipedal robot based on the balancing result and the gait data of the second bipedal robot includes: obtaining a second bipedal robot that is farthest away from the first bipedal robot in the walking direction, and performing deep learning on the secondary training model of the first bipedal robot based on the gait data of the second bipedal robot, and the learning result is used for sparse autoencoder training to obtain a four-level training model of the first bipedal robot; for the action setting of each degree of freedom of the bipedal robot, ankle joints and knee joints of the first bipedal robot, the action parameters are output by the model obtained after training the four-level training model with the gait data of the first bipedal robot.

[0038] Preferably, the first-level training model is selected by using a random selection method in a Monte Carlo search tree optimization algorithm.

[0039] Preferably, the experience data is gait data from the first bipedal robot or the second bipedal robot when it has previously passed a route that is most similar to the route information.

[0040] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A robot gait optimization method, for optimizing the gait of a first biped robot, characterized in that: The method comprises: establishing a neural network model based on a neural cell layer, training using gait data of a first biped robot obtained by a gait measurement device, and training the gait balance of the first biped robot; optimizing the gait of the first biped robot according to the balance result and the gait data of a second biped robot; The robot's feet, ankle joints and knee joints all have three degrees of freedom; The gait data includes angular velocity and acceleration, and the angular velocity and acceleration are the angular velocity and acceleration in the forward direction; The gait measurement device includes a gyroscope and an accelerometer; The training of gait data includes: Obtaining road surface information data, wherein the road surface information data at least includes route information and experience data in an electronic map that is connected to the current location; Gait data is obtained, and the pre-built neural network model is trained based on the gait data to obtain a primary training model: ， ， Where p and q represent a positive integer randomly selected between 1 and 10, and when the subscripts in the above two equations have no corresponding physical meaning, they are taken as 0; The first-level training model is used to perform deep learning on the empirical data, and the learning results are used to train the Hopfield network to obtain the second-level training model.

2. The robot gait optimization method according to claim 1, characterized in that: Gait balance training includes: Obtaining gait data of at least one second bipedal robot, wherein the number of the second bipedal robots is greater than or equal to 3; The secondary training model of the first bipedal robot is used to perform deep learning on the gait data of each second bipedal robot, and the learning results are summarized. The final learning results are used for sparse autoencoder training to obtain the third-level training model of the first bipedal robot.

3. The robot gait optimization method according to claim 2, characterized in that: Optimizing the gait of the first bipedal robot according to the balancing result and the gait data of the second bipedal robot includes: obtaining a second bipedal robot that is farthest away from the first bipedal robot in the walking direction, and performing deep learning on the second-level training model of the first bipedal robot based on the gait data of the second bipedal robot, and the learning result is used for sparse autoencoder training to obtain a four-level training model of the first bipedal robot; for the action setting of each degree of freedom of the bipedal robot, the action parameters are output by the model obtained after training the four-level training model with the gait data of the first bipedal robot.

4. The robot gait optimization method according to claim 3, characterized in that: The first-level training model is selected by using a random selection method in a Monte Carlo search tree optimization algorithm.

5. The robot gait optimization method according to claim 4, characterized in that: The experience data is gait data from the first bipedal robot or the second bipedal robot when it has previously passed a route that is most similar to the route information.

Images