Learning-Based Locomotion Controllers for Quadruped Robots in Indoor Stair Climbing via Deep Reinforcement Learning

Abstract

This paper introduces a learning-based locomotion controller for quadruped robots, employing Deep Reinforcement Learning (DRL) to enhance stair climbing capabilities in indoor settings. This approach combines a detailed locomotion controller specifically designed for navigating stairs. It also utilizes a comprehensive observation space that aids in the informed execution of terrain height information and movements. The core of this research lies in the development of a controller that achieves speeds of 1 m/s, with notable success rates of 90% and 98% for staircase configurations. This achievement demonstrates how the controller could significantly improve the adaptability and efficiency of quadruped robots in complex environments, paving the way for wider applications in challenging environments.