Real-Time Gait Adaptation for Quadrupeds

Abstract

This research presents an adaptive gait generation framework for quadrupedal robots using Model Predictive Control (MPC) integrated with Reinforcement Learning. The system enables real-time gait adaptation in unknown environments with dynamic obstacles.

Methodology

The approach combines the benefits of both learning-based and optimization-based methods:

• MPC: Provides optimal foot placement and force distribution over a prediction horizon.
• RL Fine-tuning: Adapts the MPC parameters based on terrain characteristics.
• Terrain Estimation: Online terrain classification using proprioceptive sensing.

Results

The system demonstrates successful locomotion on various terrains including flat ground, slopes, and stairs with up to 85% success rate in cluttered environments.