FA-RDP: A Frequency-Adaptive Reactive Diffusion Policy for Contact-Rich Manipulation

Lifeng Zhuo and collaborators proposed FA-RDP, a frequency-adaptive diffusion policy for contact-rich robot manipulation.

Before contact, FA-RDP uses multi-step low-frequency sampling to preserve diverse trajectories; after contact, it switches to one-step high-frequency sampling for fast reactions.

The method also adds Manifold Consistency Distillation to keep actions on the robot action manifold.

Across three tasks, FA-RDP improved success rates while preserving pre-contact trajectory diversity.