πR² · Jul 28, 2026
πR²: Reactive Real-time Flow Policies
The paper 'πR²: Reactive Real-time Flow Policies' was published on arXiv on 2026-07-28. It addresses the lack of reactivity in generalist manipulation policies that use action-chunking flow policies with large pretrained backbones. The proposed method splits conditioning into a fast channel (proprioception) and a slow channel (vision-language features), and uses a latency-adaptive flow schedule to enable reactive, real-time control while retaining large backbones and multi-modal policies.
What happened
Generalist manipulation policies often use action-chunking flow policies built on large pretrained backbones, but these run open-loop and cannot react to mid-execution sensory input. Replanning more frequently is limited by the slow perception-to-action pipeline. πR² introduces a method to make these policies reactive and real-time by splitting conditioning into a fast proprioception channel and a slow vision-language channel, and using a latency-adaptive flow schedule that treats in-flight actions as inpainting conditioning, emitting actions in one denoising step per tick.
Technical significance
The key technical contribution is the use of a per-position noise schedule from diffusion forcing, combined with a split conditioning architecture and a latency-adaptive flow schedule. This allows the policy to react to fast-changing proprioceptive data within an action chunk while tolerating stale vision-language features, and to generate actions efficiently with minimal denoising steps.
Industry impact
This work could enable more responsive and robust robotic manipulation in dynamic environments, making large-scale pretrained models practical for real-time control. It may accelerate the deployment of generalist robots in manufacturing, logistics, and service industries where closed-loop reactivity is critical.
What to watch
Next signals to watch include experimental validation on physical robots, comparisons with other reactive control methods, and integration with existing robot learning frameworks. Potential extensions could involve adaptive scheduling of vision updates or application to other sensory modalities.
Decision value
By enabling reactive real-time control with large pretrained models, πR² could reduce the engineering effort needed to deploy generalist robots in dynamic settings, potentially lowering costs and time-to-market for robotic automation solutions.