Paper Overview
Field: Machine Learning Authors: Sungjae Park, Shubham Tulsiani arXiv: 2607.26055
Abstract (translated from the Chinese forum summary)
Generalist manipulation policies increasingly take the form of action-chunking flow policies built on large pretrained backbones. Such chunks run open-loop, so the policy cannot react to sensory input arriving mid-execution, sacrificing reactivity. Replanning more often would restore it, but the perception-to-action pipeline (a large backbone plus multiple denoising steps) is too slow: this latency forbids frequent replanning and leaves committed actions stale, making such policies ill-suited for dynamic, closed-loop control.
πR² makes these policies reactive and real-time while retaining large backbones, expressive multi-modal policies, and multi-action prediction. Built on the per-position noise schedule of diffusion forcing, πR² contributes two core ideas:
1. Fast/slow conditioning channels: conditioning is split into a fast channel (proprioception, updated every tick) and a slow channel (vision-language features, updated asynchronously), so the policy reacts to proprioception within a single chunk while tolerating stale visual information. 2. Latency-adaptive flow scheduling: in-execution actions are treated as fixing conditions, allowing a single denoising step per invocation to output actions. One trained model thus adapts to different hardware latencies.
πR² requires only minimal architectural changes and can be fine-tuned from pretrained policies. Applied to GR00T-N1.7 on a real xArm6+XHand platform, it performs closed-loop replanning roughly 4x faster than the base policy (~25Hz on an A5000 GPU), executing actions every 40ms based on fresh observations.
Results
Across simulation and real-world manipulation tasks, πR² improves success rates by 23% in simulation and 30% in the real world compared to the strongest baselines.
Original Abstract (excerpt)
> Generalist manipulation policies increasingly take the form of action-chunking flow policies built on large pretrained backbones. Such chunks run open-loop, so the policy cannot react to sensory input arriving mid-execution, sacrificing *reactivity*. Replanning more often would restore it, but the perception-to-action pipeline (a large backbone plus multiple denoising steps) is too slow: this *latency* forbids frequent replanning and leaves committed actions stale, making such policies ill-suited for dynamic, closed-loop control. We present πR², which makes these policies reactive and real-time while retaining large backbones, expressive multi-modal policies, and multi-action prediction. Built on the per-position noise schedule of diffusion forcing, πR² co...
--- *Auto-collected on 2026-07-30*