NVIDIA TacitVLA cuts vision-language-action latency through reasoning/action decoupling

NVIDIA researchers report a VLA approach that uses reasoning supervision during training without delaying action inference, reducing reported latency from 495 ms to 147 ms, or up to 3.4x, while remaining competitive on RoboMME, RMBench, and memory tasks. The paper and project artifacts make this a strong efficient-embodied-AI story, though real-world generalization and independent validation remain pending.