Diffusion Policy: Visuomotor Policy Learning via Action Diffusion

123 · arXiv · GitHub

Motivation

Action representations:

• Explicit policy
  • Regression
  • Mixtures of Gaussians
  • Categorical representations of discrete actions
• Implicit policy
  • energy-based
• Diffusion policy
  • switch from explicit to implicit to better capture multi-modal distributions.

=> Reformulate robot policy as Denoising Diffusion Probabilistic Model (DDPM)

Contribution

Method

Experiment

Test content.

References

Question