Kinematic MeanFlow: One-Step Action Generation Policy for Robotic Foundation Models

Kinematic MeanFlow. One-Step Action Generation Policy for Robotic Foundation Models. By Jiawei Fan and three others, from Intel Labs China and Midea AI Research. Published on the first of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Robotic Foundation Models, or RFMs, often generate actions with multi-step flow matching, and those repeated steps make inference slow. The goal here is one-step action generation. MeanFlow offers a promising framework for exactly that, but applying it directly to RFMs makes performance collapse. The authors trace the collapse to two distinctive dynamics in the RFM velocity field. First, the local acceleration is stable early on, then surges sharply toward the end of the denoising process. Second, the spread of its magnitudes across samples widens as denoising goes on. A formulation that treats the whole process uniformly is caught out by both.

Kinematic MeanFlow, or K-MF, is a one-step action policy tailored to RFMs. It is grounded in a kinematic identity. Using that identity, K-MF splits the time derivative term in the MeanFlow formulation into two terms, one for each of two sub-intervals separated by an intermediate point.

Splitting the term lets the two halves do different jobs. One captures the early-stage denoising dynamics, where things are stable. The other captures the late-stage dynamics, where acceleration surges. The decoupled formulation also mitigates the way errors are amplified across the process.

With K-MF, RFMs achieve one-step action generation both when trained from scratch and when fine-tuned, across diverse tasks, and outperform multi-step flow matching in most settings. The speed gains are measured on GR00T-N1.6, running on an L40 and on a Jetson Orin, in both eager and compiled modes. Latency of the action head drops by between sixty-seven point five and seventy-four point four percent.

End to end, latency falls by between thirty point three and fifty-four point nine percent. The code is to be released.

The fix is precise about where MeanFlow broke. Its trouble was a velocity field that behaves one way early and another way late, with a sharp change near the end. Giving each stage its own term, joined at an intermediate point, is what turns a collapse into a one-step policy that keeps up with, and mostly beats, the multi-step one it replaces.

The two dynamics the authors found map onto the two terms. The surge in local acceleration near the end is a late-stage behaviour, and one term is given to it. The stable early behaviour gets the other. Alongside that split, the decoupled formulation is described as mitigating the amplification of error across the denoising process. The result works whether a model is trained from scratch or fine-tuned, and the latency savings appear on both the L40 and the Jetson Orin, in eager mode and compiled mode alike.