An Empirical Study of Policy Interpolation via Diffusion Models

Diffusion-based policies have shown great potential in multi-task settings, as they can solve new tasks without additional training through inference-time steering. In this paper, we explore the inference-time composition of diffusion-based policies using various interpolation methods. Our results show that, while existing methods merely switch between predefined action modes, our proposed approach can generate entirely new action patterns by leveraging existing policies, all without the need for further training or tuning.