AI

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

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 07/10/2026, 15:18:24 UTC.

Briefings

This research presents Kinematic MeanFlow (K-MF), a policy for Robotic Foundation Models that achieves one-step action generation by decoupling time derivative terms in the MeanFlow formulation. The study identifies that MeanFlow's direct application causes performance collapse due to unstable local acceleration and widening magnitude spread during denoising. K-MF mitigates these dynamics, enabling successful training and fine-tuning across diverse tasks while significantly reducing inference latency by 30.3% to 54.9% on NVIDIA Jetson Orin and L40 chips.

Key points

• Kinematic MeanFlow (K-MF) decouples time derivative terms in MeanFlow into two sub-intervals to stabilize early-stage and late-stage denoising dynamics.
• K-MF reduces action-head latency of GR00T-N1.6 by 67.5% to 74.4% across L40 and Jetson Orin in both eager and compiled modes.

Context

Robotic Foundation Models face high inference latency from multi-step flow matching. MeanFlow offers a promising framework but fails in direct application due to specific velocity field dynamics. K-MF solves this by addressing local acceleration instability and magnitude spread, enabling one-step action generation in both training and fine-tuning scenarios.

Limitations

Performance collapse in direct MeanFlow application stems from two specific velocity field dynamics: local acceleration stability early on but surging sharply at the end, and widening magnitude spread across samples during denoising. The study covers diverse tasks but does not report results on specific hardware beyond NVIDIA Jetson Orin and L40 chips.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 29/10/2026, 10:28:11 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeia2kazucylkfpd7dxxeyn4hgltrb4dvryliottbc77mczs5ckvy7e