$π\mathbf{R}^2$: Reactive Real-time Flow Policies
The paper’s core claim is that action-chunking robot policies can be made closed-loop without giving up large backbones.
πR² separates fast proprioceptive conditioning from slower vision-language features, letting the robot react inside an action chunk while vision updates asynchronously. It also uses a latency-adaptive flow schedule that treats in-flight actions as conditioning and emits actions with one denoising step per call. Finetuned from GR00T-N1.7, it replans at about 25Hz on an A5000 GPU and acts on fresh observations every 40ms. The authors report success-rate gains of up to 23% in simulation and 30% on real manipulation tasks over their strongest baseline. ArXiv · AI/CL/LG's note
πR² separates fast proprioceptive conditioning from slower vision-language features, letting the robot react inside an action chunk while vision updates asynchronously. It also uses a latency-adaptive flow schedule that treats in-flight actions as conditioning and emits actions with one denoising step per call. Finetuned from GR00T-N1.7, it replans at about 25Hz on an A5000 GPU and acts on fresh observations every 40ms. The authors report success-rate gains of up to 23% in simulation and 30% on real manipulation tasks over their strongest baseline. ArXiv · AI/CL/LG's note
score 5