πR2: Reactive Real-time Flow Policies

Modern robotics foundation models (VLAs) often rely on action-chunking flow policies that execute sequences of actions open-loop. While this improves temporal consistency, it creates a fundamental trade-off: frequent replanning is necessary for reactivity in dynamic environments, but the high computational cost of large vision-language backbones and multi-step denoising makes frequent replanning prohibitively slow. The authors seek to restore reactivity to these models without sacrificing the benefits of large-scale pretrained backbones.

To solve the latency-reactivity bottleneck, the authors propose two primary innovations. First, they disentangle the policy's conditioning into a 'fast channel' (proprioception) and a 'slow channel' (vision-language features). Proprioception is processed at every control tick, allowing the policy to react to immediate physical feedback, while the computationally expensive vision-language features are updated asynchronously. Second, they implement a latency-adaptive flow schedule based on diffusion forcing. This schedule treats in-flight actions as inpainting conditioning and reduces the inference cost to a single denoising step per call, allowing the model to adapt to varying hardware latencies and maintain smooth, real-time execution.

By applying these techniques to the GR00T-N1.7 model, the authors demonstrate that pi-R2 can replan closed-loop approximately 4x faster than standard base policies, achieving 25Hz inference on an A5000 GPU. In both simulation and real-world manipulation tasks, this increased reactivity leads to significant performance gains, improving success rates by up to 23% in simulation and 30% in real-world experiments compared to the strongest baselines. The framework is designed to be compatible with existing architectures, requiring only minimal finetuning.