πR^2: Reactive Real-time Flow Policies

Researchers introduced πR^2, a reactive real-time flow policy for robotic manipulation built on pretrained action-chunking models.

It splits fast and slow conditioning channels and uses a latency-aware one-step denoising schedule to react to fresh sensor input during execution.

In simulation and on a real xArm6+XHand setup, πR^2 replanned about four times faster than the base policy and improved task success over strong baselines.

The approach boosts closed-loop control and real-world performance without requiring a major architecture change.