πR²: Reactive Real-time Flow Policies for Robotics
A novel approach named πR² (Reactive Real-time Flow Policies) tackles the latency issue present in action-chunking flow policies utilized for generalist robot manipulation. These policies, which rely on extensive pretrained backbones, perform action chunks in an open-loop manner, thus failing to respond to sensory input during execution. Although frequent replanning could enhance reactivity, it is hindered by the slow perception-to-action pipeline, which involves a substantial backbone and multiple denoising stages. πR² enables these policies to be both reactive and real-time while maintaining large backbones, expressive multi-modal policies, and multi-action predictions. It leverages the per-position noise schedule of diffusion forcing and introduces two concepts: a fast channel for proprioception and a slow channel updated asynchronously. This research is documented in the paper "πR²: Reactive Real-time Flow Policies" on arXiv (ID: 2607.26055).
Key facts
- πR² stands for Reactive Real-time Flow Policies.
- It addresses latency in action-chunking flow policies for robot manipulation.
- The method builds on diffusion forcing's per-position noise schedule.
- It splits conditioning into fast and slow channels.
- The fast channel handles proprioception updated every tick.
- The slow channel is updated asynchronously.
- The paper is on arXiv with ID 2607.26055.
- The approach retains large pretrained backbones.
Entities
Institutions
- arXiv