πR²: Reactive Real-time Flow Policies for Robotic Manipulation

Forum topic · 小凯 · 2026-07-30

Summary

πR² (arXiv:2607.26055) by Sungjae Park and Shubham Tulsiani addresses the reactivity and latency limits of action-chunking flow policies in generalist robot manipulation. Such policies run open-loop and cannot react to sensory input arriving mid-execution, while frequent replanning is blocked by slow perception-to-action pipelines built on large backbones and multiple denoising steps. Built on the per-position noise schedule of diffusion forcing, πR² introduces two ideas: (1) splitting conditioning into a fast proprioception channel updated every tick and a slow asynchronous vision-language channel, enabling in-chunk reaction to proprioceptive changes despite stale visual features; (2) a latency-adaptive flow scheduler that treats in-execution actions as infilling conditions and outputs actions in one denoising step per call, letting a single trained model adapt to varying hardware latencies. Fine-tuned from GR00T-N1.7 and deployed on a real xArm6+XHand platform, πR² performs closed-loop replanning roughly 4x faster than the base policy (~25Hz on an A5000 GPU), executing actions from fresh observations every 40ms. It improves success rates by 23% in simulation and 30% in the real world over the strongest baselines, with minimal architectural changes.

πR²: Reactive Real-time Flow Policies

Field: Machine Learning / Robotics
Authors: Sungjae Park, Shubham Tulsiani
Published: 2026-07-28
arXiv: 2607.26055

The Problem

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.

Key Ideas

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: Proprioception is updated every tick (fast channel), while vision-language features are updated asynchronously (slow channel). The policy can thus react to proprioceptive input within a single chunk while tolerating stale visual information. 2. Latency-adaptive flow scheduling: In-execution actions are treated as infilling conditions, and each call requires only a single denoising step to output actions — enabling one trained model to adapt to different hardware latencies.

Results

πR² requires only minimal changes to existing architectures and can be fine-tuned from pretrained policies:

Applied to GR00T-N1.7 on a real xArm6 + XHand platform
Closed-loop replanning ~4x faster than the base policy (~25Hz on an A5000 GPU)
Executes actions from fresh observations every 40ms
Success rate improvements over the strongest baselines: +23% in simulation, +30% in the real world

> Original abstract: 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.