Gait in Eight: Efficient On-Robot Learning for
Omnidirectional Quadruped Locomotion
1Technical University of Darmstadt
2MAB Robotics
Conference: TBA
TLDR: We propose an on-robot learning method with joint target and CPG control architectures to achieve omnidirectional quadruped locomotion in a few minutes of training.
Overview
Our Reinforcement Learning framework enables learning on-robot locomotion in just 8 minutes of raw training time. By leveraging the efficient CrossQ algorithm along with two control architectures—Joint Target Prediction (JTP) for agile maneuvers and Central Pattern Generators (CPG) for stable, natural gaits—our approach demonstrates robustness in various indoor and outdoor environments.
Indoor Forward Locomotion
Training the CPG and JTP on the forward locomotion task in an indoor office environment. This controlled setting allows our framework to rapidly adapt and refine the robot's gait with minimal external disturbances. The JTP architecture reached forward velocities of up to 0.85 m/s.
Outdoor Forward Locomotion
Training the CPG and JTP on the forward locomotion task in an outdoor pavement environment. Despite the challenges of natural terrain, our approach consistently achieves stable and efficient forward motion with both control architectures.
Outdoor Omnidirectional Locomotion
Training the CPG and JTP on the omnidirectional locomotion task in an outdoor pavement environment. This setup showcases the framework's ability to handle goal-conditioned multidirectional movement and adapt to rough real-world conditions.
Acknowledgments
This project was funded by National Science Centre, Poland, under the OPUS call in the Weave program UMO-2021/43/I/ST6/02711, by the German Science Foundation (DFG) under grant number PE 2315/17-1 and by the project ``Third Wave of AI'', funded by the Excellence Program of the Hessian Ministry of Higher Education, Science, Research and Art.
This website was inspired by Kevin Zakka's and Brent Yi's and builds on Nico Bohlinger's.
Code
Paper
Video