Intel's Kinematic MeanFlow Cuts GR00T Action-Head Latency by up to 74% with One-Step Generation
IntelLabsChina · hf · 2026-10-07
Intel Labs China introduces Kinematic MeanFlow (K-MF), a one-step action generation policy for robotic foundation models. Direct MeanFlow application collapses because RFM velocity fields show late-stage acceleration surges and widening magnitude spread; K-MF decouples the time-derivative term into two sub-interval terms around an intermediate point to capture early and late denoising dynamics. It enables one-step generation from scratch or fine-tuning, outperforming multi-step flow matching in most tasks, and cuts GR00T-N1.6 action-head latency by 67.5%74.4% on L40 and Jetson Orin (30.3%54.9% end-to-end). Code will be open-sourced.
More from Embodied
- Robotics company LeRobot jokes it now makes music videos too — Thom_Wolf · 2026-10-08
- MIT's SOLE-R1: Video-Language Reasoning as the Sole Reward Enables Zero-Shot On-Robot RL — micoolcho · 2026-10-08
- 200 builders revive vintage toys as voice agents at AssemblyAI hackathon — AssemblyAI · 2026-10-07
- Robotics is having its moment: why roboticists are Renaissance generalists — lukas_m_ziegler · 2026-10-07
- PhD student wraps up robotics planning series with EMPIRIC residual world model — tomssilver · 2026-10-07
- Wearing a BCI headband like Chinese quant traders is the new productivity hack — alexbilz · 2026-10-07