Kinematic MeanFlow: One-Step Action Generation Policy for Robotic Foundation Models
K-MF is pitched as a one-step action policy that cuts robotic action-head latency by 67.5% to 74.4% on GR00T-N1.6.
The paper says direct MeanFlow use in robotic foundation models collapses because late-stage denoising dynamics become unstable and uneven across samples. Kinematic MeanFlow splits the time-derivative term around an intermediate point so early and late denoising behavior can be handled separately. The authors report one-step action generation across training-from-scratch and fine-tuning settings, outperforming multi-step flow matching in most tested cases. End-to-end latency drops are reported at 30.3% to 54.9%. HF Daily Papers' note
The paper says direct MeanFlow use in robotic foundation models collapses because late-stage denoising dynamics become unstable and uneven across samples. Kinematic MeanFlow splits the time-derivative term around an intermediate point so early and late denoising behavior can be handled separately. The authors report one-step action generation across training-from-scratch and fine-tuning settings, outperforming multi-step flow matching in most tested cases. End-to-end latency drops are reported at 30.3% to 54.9%. HF Daily Papers' note
score 4