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

FA-RDP realizes frequency-adaptive visual-force control through a shared multi-frequency visual-force Transformer. The same backbone is used for both frequency modes: a single forward pass is conditioned on one frequency mode $\nu \in \{\ell, h\}$ and predicts the action chunk for that mode.

The key design is a frequency-adaptive positional encoding. The frequency mode is not embedded as an extra token; instead, it selects temporal indices on a shared temporal grid. The high-frequency mode uses consecutive positions, whereas the low-frequency mode uses sparse positions, so both modes are expressed on the same underlying clock. We use a causal action-force mask to avoid future-contact leakage.

The multi-frequency Transformer provides both low-frequency and high-frequency diffusion samplers, but the policy still needs to decide which sampler to use at different phases. We use a learned indicator that estimates local multimodality instead of relying on a phase label.

At inference time, FA-RDP computes the indicator value from the current slow tokens and selects the sampler with a fixed threshold. This makes execution frequency-adaptive, but the selected high-frequency diffusion sampler still requires multi-step denoising. We therefore distill the high-frequency sampler for fast force-feedback response.

Manifold consistency distillation (MCD) turns the selected high-frequency sampler into a one-step sampler. Directly distilling a model that predicts epsilon, score, or velocity requires the student to fit high-frequency, noise-like targets, which makes stable prediction difficult.

Inspired by the sample-prediction view of JiT and Pixel Mean Flows, we reparameterize the diffusion model to predict action chunks on the robot action manifold. The fixed DDPM conversion retains residual supervision without making the network predict a noise-like target, reducing the student's learning difficulty. After distillation, FA-RDP uses the multi-frequency checkpoint for the low-frequency sampler and the distilled checkpoint for the high-frequency sampler.