MorphDLoco: Morphology aware Diffusion policy for multi-robot multi-skill locomotion

Learning from offline datasets poses scalability challenges for robotics due to their limited availability. Furthermore, current offline learning techniques in robotics either lack the ability to translate effectively to real-time deployment or fail to capture the inherent diversity within the datasets. In this work, we propose a diffusion learning framework as a unified policy capable of distilling multiple skills across a variety of robotic platforms into a single architecture. We leverage diffusion’s multimodal capabilities to learn from diverse datasets, demonstrating that our single diffusion-based policy can learn multiple locomotion gaits—such as trot, pronk, bound, and pacing—across quadrupedal robots with varying morphologies and physical parameters. An emergent advantage of our method is its ability to transfer learned skills to a robot with a new morphology in a zero-shot manner.