Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning Multiple Gaits within Latent Space for Quadruped Robots

Jinze Wu, Yufei Xue, Chenkun Qi

TL;DR

This work designs gait-dependent rewards that are constructed explicitly from gait parameters and implicitly from conditional adversarial motion priors (CAMP), and demonstrates such multiple gaits control on a quadruped robot Go1 with only proprioceptive sensors.

Abstract

Learning multiple gaits is non-trivial for legged robots, especially when encountering different terrains and velocity commands. In this work, we present an end-to-end training framework for learning multiple gaits for quadruped robots, tailored to the needs of robust locomotion, agile locomotion, and user's commands. A latent space is constructed concurrently by a gait encoder and a gait generator, which helps the agent to reuse multiple gait skills to achieve adaptive gait behaviors. To learn natural behaviors for multiple gaits, we design gait-dependent rewards that are constructed explicitly from gait parameters and implicitly from conditional adversarial motion priors (CAMP). We demonstrate such multiple gaits control on a quadruped robot Go1 with only proprioceptive sensors.

Learning Multiple Gaits within Latent Space for Quadruped Robots

TL;DR

This work designs gait-dependent rewards that are constructed explicitly from gait parameters and implicitly from conditional adversarial motion priors (CAMP), and demonstrates such multiple gaits control on a quadruped robot Go1 with only proprioceptive sensors.

Abstract

Learning multiple gaits is non-trivial for legged robots, especially when encountering different terrains and velocity commands. In this work, we present an end-to-end training framework for learning multiple gaits for quadruped robots, tailored to the needs of robust locomotion, agile locomotion, and user's commands. A latent space is constructed concurrently by a gait encoder and a gait generator, which helps the agent to reuse multiple gait skills to achieve adaptive gait behaviors. To learn natural behaviors for multiple gaits, we design gait-dependent rewards that are constructed explicitly from gait parameters and implicitly from conditional adversarial motion priors (CAMP). We demonstrate such multiple gaits control on a quadruped robot Go1 with only proprioceptive sensors.
Paper Structure (16 sections, 8 equations, 5 figures, 5 tables)

This paper contains 16 sections, 8 equations, 5 figures, 5 tables.

Table of Contents

  1. INTRODUCTION
  2. Related RL Methods for Locomotion
  3. Contributions
  4. METHOD
  5. Reinforcement Learning Problem Formulation
  6. Network Architecture
  7. Latent Space for Gait Skills
  8. Reward Functions Design
  9. Training
  10. Training Setup
  11. Training Curriculum
  12. RESULTS AND DISCUSSION
  13. Latent Space Analysis
  14. Ablation Study for the Learning Framework
  15. Outdoor Experiments
  16. ...and 1 more sections

Figures (5)

  • Figure 1: Overview of the learning framework. We employ the asymmetric actor-critic framework to learn robust blind locomotion in one training phase. A latent space $\boldsymbol{Z}$ is constructed concurrently by a gait encoder and a gait generator, which helps the agent to reuse multiple gait skills to achieve adaptive gait behaviors. We use a conditional discriminator to guide the policy to reproduce the natural gait motions that are similar to the CAMP dataset generated from TO.
  • Figure 2: The original space of the gait parameters (left) and the latent space of the gait skills (right).
  • Figure 3: The t-SNE visualization for the gait skills in the latent space.
  • Figure 4: The DTW distance of the gait skills in the latent space. Darker colors represent higher proximity between the two gait skills in the latent space.
  • Figure 5: The robust (left) and agile (right) tests outdoors.