One-Step Flow Policy: Self-Distillation for Fast Visuomotor Policies

A team of researchers has introduced the One-Step Flow Policy (OFP), a novel framework designed to solve a critical bottleneck in AI-powered robotics. Current high-precision robotic policies often rely on generative flow or diffusion models, which produce excellent, multimodal action distributions but require slow, iterative sampling. This latency severely degrades control frequency, making real-time, time-sensitive manipulation difficult. OFP addresses this by employing a from-scratch self-distillation approach, meaning it trains a fast, single-step policy without needing a pre-trained, slow teacher model. It combines a self-consistency loss to ensure coherent predictions across time and a self-guided regularization to sharpen actions toward proven, high-performance modes.

In comprehensive evaluations, OFP demonstrated its practical superiority. Tested across 56 diverse simulated manipulation tasks, a one-step OFP achieved state-of-the-art results, outperforming policies that required 100 iterative steps. This translated to an action generation speed-up of over 100 times. The researchers further validated OFP's real-world potential by integrating it into the π₀.₅ model on the RoboTwin 2.0 platform, where the one-step version surpassed the original 10-step policy's performance. These results position OFP as a scalable solution for deploying highly accurate, low-latency AI controllers in dynamic physical environments, bridging the gap between complex AI reasoning and the hard real-time requirements of robotics.