Action Diffusion

Generate robot action sequences using diffusion models, enabling smooth, multimodal, and temporally consistent behavior generation for complex manipulation and locomotion tasks. Action diffusion treats action generation as a denoising process: starting from random noise in action space, a diffusion model iteratively denoises toward a valid action sequence. The model is conditioned on visual observations and (optionally) goal states. The key advantages over direct regression are multimodality (multiple valid actions for the same observation) and temporal consistency (actions are generated as a sequence, not per-timestep). This produces smoother, more robust robot behavior.