Researchers have proposed Kinematic MeanFlow (K-MF), a one-step action generation policy designed for Robotic Foundation Models (RFMs). The work targets the high inference latency caused by multi-step flow matching, a common bottleneck in deploying such models.
According to the paper, MeanFlow offers a promising route to one-step generation, but applying it directly to RFMs causes performance collapse. The authors trace this to two dynamics in the RFM velocity field: local acceleration stays stable early in denoising but surges sharply near the end, and the spread of its magnitudes across samples widens as denoising progresses.
K-MF addresses these issues by using a kinematic identity to decouple the time derivative term in the MeanFlow formulation into two sub-interval terms separated by an intermediate point. This lets each term capture early-stage and late-stage denoising dynamics respectively, while mitigating error amplification across the process.
The approach enables RFMs to perform one-step action generation in both training-from-scratch and fine-tuning paradigms across diverse tasks, and the authors report it outperforms multi-step flow matching in most settings. On inference efficiency, K-MF reduces the action-head latency of GR00T-N1.6 by 67.5%–74.4% across L40 and Jetson Orin hardware in eager and compiled modes, translating to end-to-end latency reductions of 30.3%–54.9%.
The authors state that code will be available at https://github.com/IntelChina-AI/K-MF.