Locomotion Policy Learning via Diffusion Policy

The emergence of deep reinforcement learning has recently led to remarkable achievements in legged locomotion. Compared to traditional model-based approaches, reinforcement learning-based control methods can improve robustness and generalization in the face of environmental uncertainties. However, due to the complexity of the locomotion policy, the learned gaits are generally conservative and lack naturalness. In this paper, we propose a novel framework for learning locomotion policy that results in gaits characterized by both robustness and generalization. We incorporate a diffusion model into our policy learning framework for legged locomotion. The diffusion model powerfully represents policy, leading to multimodal action distributions and sufficient exploration.