piR2 Splits Fast Proprioception From Slow Vision to Make Flow Policies Replan 4x Faster at ~25Hz

arXiv 2607.26055·medium signalXBlueskyLinkedInCopy link

Posted 2026-07-28 (arXiv 2607.26055), piR2 addresses why action-chunking flow policies built on large pretrained backbones cannot do closed-loop control: chunks run open-loop so mid-execution sensory input is ignored, and the perception-to-action pipeline is too slow to replan often enough. Building on diffusion forcing's per-position noise schedule, it splits conditioning into a fast channel (proprioception, refreshed every tick) and an asynchronously updated slow channel (vision-language features), and adds a latency-adaptive flow schedule that treats in-flight actions as inpainting conditioning and emits actions in one denoising step per call, so one trained model adapts to varying hardware latency. Applied to GR00T-N1.7 on a real xArm6+XHand platform it replans roughly 4x faster (~25Hz on an A5000, fresh observation every 40ms) and raises success up to 23% in simulation and 30% in the real world.