$\pi\mathbf{R}^2$: Reactive Real-time Flow Policies

Researchers introduced piR2, a reactive flow-policy framework for robot manipulation that separates fast proprioceptive input from slower vision-language features.

piR2 enables frequent, latency-aware replanning with a one-step denoising schedule and can be finetuned from pretrained models like GR00T-N1.7.

The system runs at about 25 Hz on an A5000 GPU with fresh observations every 40 ms, making closed-loop control more responsive to new inputs.

It improves success rates over the strongest baseline in both simulation and real-world manipulation tasks with xArm6 and XHand.