A new technique lets robot AI models generate actions in one step instead of many, without the usual loss in quality.
Researchers built Kinematic MeanFlow (K-MF), a training method that lets Robotic Foundation Models produce a full action in a single pass instead of running through dozens of denoising steps. Standard flow-matching models need those extra steps to work well, and earlier attempts to compress the process into one step caused performance to collapse. The team traced the failure to the model's internal velocity field: it stays stable early in denoising, then spikes sharply near the end, with results varying more from sample to sample as the process continues. K-MF splits the underlying math at an intermediate point so the early and late dynamics are handled separately, which keeps the one-step version from falling apart.
The payoff shows up as real hardware speed, not just a cleaner equation. On GR00T-N1.6, a widely referenced robotics foundation model, K-MF cut action-head latency by up to 74% on L40 GPUs and Jetson Orin edge chips, and trimmed total end-to-end latency by up to 54.9%. That matters because a robot moving in the physical world can't wait half a second for a model to finish thinking.
Edge inference has been the quiet bottleneck in robotics AI while the industry argues about model size. The code is promised but not yet public, and these numbers come from benchmarks, not a robot handling dishes in a messy kitchen.