Nvidia's TacitVLA cuts reasoning VLA latency up to 3.4x with no performance loss
nvidia · hf · 2026-10-14
Nvidia introduces TacitVLA, a reasoning vision-language-action model that avoids reasoning latency. Key insight: actions need not wait for or condition on autoregressively decoded reasoning text, since reasoning supervision already improves backbone representations. The action expert (flow matching) never attends to the reasoning generated for the current call; on long-horizon tasks the reasoning text doubles as episodic memory and entries are generated during execution.
Controlled studies across four benchmarks spanning manipulation and autonomous driving (RoboMME, RMBench, LIBERO, PhysicalAI-AV) show:
- Performance matching its explicit-reasoning counterpart at up to 3.4x lower latency;
- Up to +17.1% success over the same architecture trained without reasoning;
- New best published results on RoboMME Counting (82.7% vs 66.8%) and RMBench (77.5% vs 56.5%), plus 72.9% on LIBERO-Plus and a 52.5% scene score on the PhysicalAI-AV closed-loop driving challenge.
Project page: tacitvla.github.io
More from Embodied
- Robotics labs converge on RL experts distilled into a generalizable controller — KyleMorgenstein · 2026-10-14
- AI glasses ship with zero third-party apps; phone projection is the only workaround — alexisgallagher · 2026-10-14
- AMD robotics director interviews Foundation CEO on humanoid robots for heavy industry — sankaet · 2026-10-14
- EgoExplorer: 100-hour vision-motion-touch dataset for humanoid loco-manipulation — InternRobotics · 2026-10-14
- Guy jokes a robot pulled more women at the bar than he did — jonstephens85 · 2026-10-14
- NYU's Lerrel Pinto: many robot types now share one intelligence layer, thanks to PI — svlevine · 2026-10-14