Researchers have built a robot-control AI that needs just two passes through its neural network to decide on an action, instead of the dozens most systems require.
The paper, called HybridFlow, splits inference into three stages that all reuse one trained model. A "Global Jump" step uses a technique called MeanFlow to sketch a rough action trajectory in one shot. A parameter-free step called ReNoise then builds an intermediate state partway through that trajectory. A final "Local Refine" step queries the same network again to sharpen the result, with no separate distillation model required. In simulated RoboMimic manipulation tests, this two-step recipe hit 95-95.5% success, compared with 78% for a single-step version of MeanFlow alone. Across five real-robot tasks, all running on the same Jetson AGX Thor hardware, HybridFlow beat a 16-step Diffusion Policy baseline by 13 to 68 points on normalized performance while generating actions roughly eight times faster.
That speed matters because diffusion-based robot policies have been accurate but slow, iterating through dozens of denoising steps that make real-time control on embedded hardware, the kind actually bolted to a robot arm, a genuine bottleneck. Getting comparable or better accuracy in two steps, and showing it also works as an "action expert" plugged into a vision-language-action system, points toward faster embodied-AI stacks without the usual latency tax.
Still, this is one arXiv paper with self-reported benchmarks and a five-task real-robot test, not an independently verified industry standard, so treat the eightfold speedup as a lab result until other teams reproduce it.