πR^2: Reactive Real-time Flow Policies
The paper’s claim is a faster closed-loop robot policy that can keep reacting during action chunks.
πR² splits fresh proprioception from slower vision-language features, then uses a latency-adaptive flow schedule to update actions with one denoising step per call. The authors say it can be fine-tuned from existing architectures and tested it on GR00T-N1.7 with an xArm6+XHand setup. It replans at about 25Hz on an A5000 GPU, acting on a fresh observation every 40ms. Reported gains reach up to 23% in simulation and 30% in real-world manipulation over the strongest baseline. HF Daily Papers' note
πR² splits fresh proprioception from slower vision-language features, then uses a latency-adaptive flow schedule to update actions with one denoising step per call. The authors say it can be fine-tuned from existing architectures and tested it on GR00T-N1.7 with an xArm6+XHand setup. It replans at about 25Hz on an A5000 GPU, acting on a fresh observation every 40ms. Reported gains reach up to 23% in simulation and 30% in real-world manipulation over the strongest baseline. HF Daily Papers' note
score 5