FLASH Policy: 175x faster robot control with single-step inference

A team of researchers (Jiaqi Bai, Jindou Jia, Yuxuan Hu, Gen Li, Xiangyu Chen, Tuo An, Kuangji Zuo, Jianfei Yang) has published FLASH (Fast Legendre-polynomial Action policy via Sparse History-anchored flow) on arXiv, a new approach to visuomotor policy learning for robotics. The key innovation is replacing the common two-step process of generating discrete action chunks and then denoising them iteratively. Instead, FLASH represents robot actions as continuous Legendre polynomial trajectories derived from expert demonstrations under sparse temporal sampling. This allows a single forward pass of a flow-matching model to cover an extended action horizon, completely sidestepping the slow denoising loop required by diffusion or prior flow-matching policies.

The performance gains are dramatic. In experiments across five simulated and two real-world manipulation tasks, FLASH achieves state-of-the-art success rates of at least 92% across all tasks. The per-episode inference time is just 31.40 milliseconds — that’s up to 175 times faster than diffusion policies and 18 times faster than previous flow-matching methods. Training converges up to 4 times faster than the popular ACT (Action Chunking with Transformers) approach. Furthermore, because FLASH outputs continuous trajectories, the analytic differentiation of the polynomial directly provides velocity feed-forward signals to torque controllers, reducing tracking error by 5x to 7x compared to discrete-action baselines. This work points toward a future where real-time, reactive robotic control can leverage generative models without the latency penalty that has limited their deployment.