CN-121477655-B - Horse-step generation and control method, device, equipment and medium for four-foot machine

CN121477655BCN 121477655 BCN121477655 BCN 121477655BCN-121477655-B

Abstract

The invention provides a four-foot machine horse gait generating and controlling method, device, equipment and medium, which comprise the steps of constructing gait animation data of a virtual horse, redirecting the gait animation data of the virtual horse into a kinematic model of the four-foot machine horse according to a mapping relation to obtain reference track data, constructing a target data set according to the reference track data, carrying out parameterized track expansion on the target data set through predefined continuously adjustable parameters, pre-training a machine horse control strategy according to the target data set through countermeasure type imitation learning, carrying out reinforcement learning joint training on the machine horse control strategy in a physical simulation environment through predefined high-level parameterized instructions and task rewarding mechanisms to obtain a unified gait generating strategy, and controlling the four-foot machine horse to execute corresponding actions through the unified gait generating strategy. Thereby realizing the high-efficiency control of the shape of the quadruped machine horse in the real environment.

Inventors

Request for anonymity
Request for anonymity

Assignees

杭州云深处科技股份有限公司

Dates

Publication Date: 20260512
Application Date: 20260109

Claims (9)

1. A horse step generation and control method of a four-foot machine is characterized by comprising the following steps: constructing gait animation data of a virtual horse, wherein the gait animation data of the virtual horse comprises a key frame animation corresponding to a preset specific gait of the virtual horse and a key frame animation corresponding to a preset specific behavior; Aligning the skeleton model of the virtual horse with the joint topology of the quadruped robot horse, and establishing a mapping relation between the animation skeleton joint and the robot horse joint; Redirecting the gait animation data of the virtual horse into a kinematic model of a quadruped robot horse according to the mapping relation between the animation skeleton joints and the robot horse joints to obtain executable reference track data of the quadruped robot horse in a preset joint range; constructing a target data set for countering imitation learning according to the reference track data; constructing predefined continuously adjustable parameters; The method comprises the steps of carrying out parameterization track expansion on a target data set through predefined continuous adjustable parameters, carrying out parameterization track expansion on the target data set through the predefined continuous adjustable parameters, and carrying out elastic deformation on an animation time axis of the target data set through a time scaling coefficient, a gait step frequency coefficient and a stride scaling coefficient respectively to generate reference track variants with different rates, reference track variants with different stride, and reference track variants with different stride; pre-training a machine horse control strategy according to the target data set through countermeasure imitation learning; Performing reinforcement learning combined training on the machine horse control strategy in a physical simulation environment through a predefined high-level parameterized instruction and a task rewarding mechanism to obtain a unified gait generation strategy for continuously adjusting the gait type-step frequency-speed-behavior mode; the quadruped machine horse is controlled to perform corresponding actions by the unified gait generating strategy for continuously adjusting gait type-stride frequency-speed-behavior patterns.
2. The method of claim 1, wherein redirecting the gait animation data of the virtual horse into a kinematic model of a quadruped machine horse according to a mapping relationship between the animated skeletal joints to the machine horse joints to obtain executable reference trajectory data of the quadruped machine horse corresponding to a preset joint range, comprises: acquiring gait related data from the gait animation data of the virtual horse according to the mapping relation between the animation skeleton joint and the machine horse joint; Redirecting the gait related data into a kinematic model of a quadruped machine horse, and generating initial trajectory data of the quadruped machine horse, which corresponds to executable within a preset joint range; And carrying out data integration on the initial track data according to the time sequence requirement to obtain executable reference track data of the quadruped robot corresponding to a preset joint range.
3. The method of claim 1, wherein constructing a target dataset for countering simulation learning from the reference trajectory data comprises: Performing morphological transformation on the reference track data to obtain an initial data set for resisting imitation learning; And extracting gait style related features from the initial dataset, and performing physical simulation on the gait style related features through an environmental physical engine to obtain a target dataset for counterimitative learning.
4. The method of claim 1, wherein reinforcement learning joint training of the machine horse control strategy in a physical simulation environment by predefined high-level parameterized instructions and task rewards mechanisms results in a unified gait generation strategy for continuously adjusting gait type-step frequency-speed-behavior patterns, comprising: randomly sampling different high-level parameterized instructions, executing the machine horse control strategy in a physical simulation environment, and summarizing the obtained reinforcement learning return according to the task rewarding mechanism; and carrying out strategy optimization on the machine horse control strategy according to the reinforcement learning return through a reinforcement learning algorithm to obtain the unified gait generation strategy for continuously adjusting the gait type-step frequency-speed-behavior mode.
5. The method as recited in claim 1, further comprising: The user requirements are abstracted to the high-level parameterized instructions including a desired forward speed, a desired lateral speed, a desired steering angular speed, a desired gait type, a target stride frequency, a target stride, and a target behavior pattern.
6. The method of claim 4, wherein the task rewarding mechanism is configured to describe a tracking rewards, a task rewards, a gait phase and step frequency consistency rewards, a stability and safety rewards, wherein the tracking rewards are configured to restrict the quadruped machine horse from following a reference gait generated by a combination of animation data and high-level instructions in a physical simulation, wherein the task rewards are configured to restrict the quadruped machine horse from successfully completing a specified displacement task while satisfying a preset behavioral style, wherein the gait phase and step frequency consistency rewards are configured to align a time of a foot contact/swing pattern of the quadruped machine horse with a given step frequency and phase schedule, and wherein the stability and safety rewards are configured to restrict limb stability of the quadruped machine horse while performing a corresponding action.
7. A four-legged machine horse step generation and control device, comprising: The first construction module is used for constructing gait animation data of the virtual horse, wherein the gait animation data of the virtual horse comprises a key frame animation corresponding to a preset specific gait of the virtual horse and a key frame animation corresponding to a preset specific behavior; The building module is used for aligning the skeleton model of the virtual horse with the joint topology of the quadruped robot horse and building a mapping relation between the animation skeleton joint and the robot horse joint; the redirection module is used for redirecting the gait animation data of the virtual horse into a kinematic model of a quadruped robot horse according to the mapping relation between the animation skeleton joint and the robot horse joint to obtain executable reference track data of the quadruped robot horse in a range corresponding to a preset joint; a second construction module for constructing a target data set for countering the imitation learning from the reference trajectory data; A third construction module for constructing predefined continuously adjustable parameters; The system comprises an expansion module, an animation style adjustment module, a reference track modification module and a reference track modification module, wherein the expansion module is used for carrying out parameterization track expansion on the target data set through predefined continuous adjustable parameters, and is particularly used for carrying out elastic deformation on an animation time axis of the target data set through a time scaling coefficient, a gait step frequency coefficient and a stride scaling coefficient to generate reference track modification of different speeds, reference track modification of unsynchronized frequencies and reference track modification of different strides; the first training module is used for pre-training the machine horse control strategy according to the target data set through antagonistic imitation learning; The second training module is used for performing reinforcement learning combined training on the machine horse control strategy in a physical simulation environment through a predefined high-level parameterized instruction and task rewarding mechanism to obtain a unified gait generation strategy for continuously adjusting the gait type, the step frequency, the speed and the behavior mode; the control module is used for controlling the quadruped machine to execute corresponding actions through the unified gait generating strategy for continuously adjusting the gait type-step frequency-speed-behavior mode.
8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the four-legged machine horse step generation and control method according to any one of claims 1 to 6 when executing the computer program.
9. A computer readable storage medium having a computer program stored thereon, wherein the computer program when executed by a processor implements the four-legged machine horse step generation and control method according to any one of claims 1 to 6.

Description

Horse-step generation and control method, device, equipment and medium for four-foot machine Technical Field The embodiment of the disclosure relates to the technical field of artificial intelligence, in particular to a horse step generation and control method, device, equipment and medium suitable for a four-foot machine. Background The four-foot robot horse is a bionic robot with four leg execution mechanisms, the body structure comprises a trunk, a head, four limbs and a tail, a high-torque servo motor or a hydraulic/pneumatic executor can be adopted, and the four-foot robot horse is matched with a multi-degree-of-freedom joint design so as to simulate the biological motion characteristics of equine animals. The mechanical gait of the quadruped robot horse can simulate various typical motion modes such as walking, jogging, running and the like of a real horse in a highly bionic mode, an accurate gait parameter model is constructed through collection and analysis of the kinematic data of the equine animals, and dynamic adjustment of key motion parameters such as stride, stride frequency, leg lifting height, landing angle and the like is realized. However, the gait generating efficiency in the prior art is low, and it is difficult to effectively control the movement pattern of the machine horse. Disclosure of Invention Embodiments described herein provide a four-foot machine horse step generation and control method, apparatus, device, and medium that overcome the above-described problems. In a first aspect, according to the present disclosure, there is provided a four-legged machine horse step generation and control method, including: constructing gait animation data of a virtual horse, wherein the gait animation data of the virtual horse comprises a key frame animation corresponding to a preset specific gait of the virtual horse and a key frame animation corresponding to a preset specific behavior; Aligning the skeleton model of the virtual horse with the joint topology of the quadruped robot horse, and establishing a mapping relation between the animation skeleton joint and the robot horse joint; Redirecting the gait animation data of the virtual horse into a kinematic model of a quadruped robot horse according to the mapping relation between the animation skeleton joints and the robot horse joints to obtain executable reference track data of the quadruped robot horse in a preset joint range; constructing a target data set for countering imitation learning according to the reference track data; constructing predefined continuously adjustable parameters; performing parameterized track expansion on the target data set through predefined continuously adjustable parameters; pre-training a machine horse control strategy according to the target data set through countermeasure imitation learning; Performing reinforcement learning combined training on the machine horse control strategy in a physical simulation environment through a predefined high-level parameterized instruction and a task rewarding mechanism to obtain a unified gait generation strategy for continuously adjusting the gait type-step frequency-speed-behavior mode; the quadruped machine horse is controlled to perform corresponding actions by the unified gait generating strategy for continuously adjusting gait type-stride frequency-speed-behavior patterns. In a second aspect, according to the present disclosure, there is provided a quadruped machine horse step generation and control device, comprising: The first construction module is used for constructing gait animation data of the virtual horse, wherein the gait animation data of the virtual horse comprises a key frame animation corresponding to a preset specific gait of the virtual horse and a key frame animation corresponding to a preset specific behavior; The building module is used for aligning the skeleton model of the virtual horse with the joint topology of the quadruped robot horse and building a mapping relation between the animation skeleton joint and the robot horse joint; the redirection module is used for redirecting the gait animation data of the virtual horse into a kinematic model of a quadruped robot horse according to the mapping relation between the animation skeleton joint and the robot horse joint to obtain executable reference track data of the quadruped robot horse in a range corresponding to a preset joint; a second construction module for constructing a target data set for countering the imitation learning from the reference trajectory data; A third construction module for constructing predefined continuously adjustable parameters; the expansion module is used for expanding the parameterized track of the target data set through predefined continuously adjustable parameters; the first training module is used for pre-training the machine horse control strategy according to the target data set through antagonistic imitation learning; The second training module is used for per