Learning an Adaptive Fall Recovery Controller for Quadrupeds on Complex Terrains

TL;DR

This paper introduces an adaptive fall recovery controller for quadruped robots using deep reinforcement learning, enabling effective recovery on complex terrains like rocky slopes and irregular stones, with successful transfer to real robots.

Contribution

The paper presents a novel deep RL-based AFR controller that generalizes across diverse terrains and is transferable to multiple quadrupedal platforms.

Findings

Outperforms baseline methods in success rate

Demonstrates effective transfer from simulation to real robots

Generalizes well to various complex terrains

Abstract

Legged robots have shown promise in locomotion complex environments, but recovery from falls on challenging terrains remains a significant hurdle. This paper presents an Adaptive Fall Recovery (AFR) controller for quadrupedal robots on challenging terrains such as rocky, breams, steep slopes, and irregular stones. We leverage deep reinforcement learning to train the AFR, which can adapt to a wide range of terrain geometries and physical properties. Our method demonstrates improvements over existing approaches, showing promising results in recovery scenarios on challenging terrains. We trained our method in Isaac Gym using the Go1 and directly transferred it to several mainstream quadrupedal platforms, such as Spot and ANYmal. Additionally, we validated the controller's effectiveness in Gazebo. Our results indicate that the AFR controller generalizes well to complex terrains and outperforms baseline methods in terms of success rate and recovery speed.