πR² addresses the open-loop behavior of action-chunking flow policies by separating conditioning into a fast proprioceptive channel and an asynchronously refreshed vision-language channel. Its latency-adaptive flow schedule treats in-flight actions as inpainting constraints and performs one denoising step per call, allowing replanning under variable hardware latency. Fine-tuned from GR00T-N1.7, the system reportedly reaches roughly 25 Hz on an A5000 GPU with an xArm6+XHand platform, acting on a fresh observation every 40 ms. The paper reports up to 23% higher simulation success and 30% higher real-world success than the strongest baseline.
No heat snapshots are available in the last 24 hours.