Reactive Diffusion Policy: plan ahead, respond now
What changes when planning a movement and responding to fingertip contact run at different speeds?
INTERACT ROBOTICS · Research Notes · Reviewed September 7, 2026
Contact can change mid-plan.
Policies that generate sequences of actions can produce smooth motion. But objects and contact conditions can change while those actions execute. Reactive Diffusion Policy (RDP) combines slower visual planning with fast tactile and force response.
When the disturbance happens matters.
For peeling with post-contact disturbance, the paper reports a score of 0.19 for DP and 0.88 for the force-input RDP configuration. These are task-condition performance scores, not average success across all tasks. Optical tactile and force-sensor configurations must also be distinguished.
Responsiveness is a system property.
A fast sensor alone does not make a fast response. Communication, preprocessing, inference, and robot control all add delay. The relevant measure is the full time from a contact change to an action correction.
Speed is only part of the question.
Fast corrections must not disrupt the larger plan or overreact to noise. Separating pre-contact approach from post-contact tracking makes it clearer which layer needs improvement.
References
- Reactive Diffusion Policy: Slow-Fast Visual-Tactile Policy Learning for Contact-Rich Manipulation ↗
2025 / Table II, Peeling / Post-contact disturbance. The values 0.19 and 0.88 are task-performance scores from that table.