Found first: a primary source the press has not covered yet.
Researchers from Intel Labs China and Midea AI Research have published Kinematic MeanFlow (K-MF), a one-step action-generation policy for GR00T-N1.6 that cuts action-head inference latency 67.5%-74.4% and end-to-end latency 30.3%-54.9% compared to the multi-step flow matching baseline. On most benchmark tasks the one-step policy matches or slightly exceeds multi-step performance; on one of the four LIBERO suites it falls 0.5 points short.
What the source says
K-MF applies a kinematic identity to decompose the velocity field into two sub-interval terms, targeting two failure modes the authors identify in standard flow matching: unstable local acceleration late in denoising and widening velocity magnitude spreads across samples. Latency tests run on NVIDIA L40 (desktop) and Jetson Orin (edge) in both eager and compiled execution modes. On the four LIBERO task suites fine-tuned from GR00T-N1.6, K-MF scores 99.5/99.5/98.0/94.5 versus 97.5/98.5/98.5/94.5 for multi-step flow matching. Cross-domain evaluations on Fractal (78.4% vs 73.9%), BridgeData V2 (59.9% vs 58.4%), and PointNav (74.0% vs 73.8%) show K-MF at or above multi-step baselines, with training overhead of approximately 33% more wall-clock time for GR00T-N1.6 versus flow matching and a GPU memory increase of at most 2.5% over the MeanFlow baseline. Code is published at github.com/IntelChina-AI/K-MF.
Why it matters
Multi-step flow matching's iterative denoising loop is a concrete latency bottleneck for real-time robotics, particularly on edge hardware like Jetson Orin where compute budgets are fixed. Cutting action-head latency by more than two-thirds without meaningful task-success loss changes what is practical to deploy on a physical robot. The cross-domain results across three datasets and different robot morphologies suggest the kinematic decomposition is not specific to LIBERO or to GR00T-N1.6, though the paper itself tests only two host models. The 33% additional training time for GR00T-N1.6 is the primary cost of adoption.