UnderwaterVLA separates mission reasoning from reactive vehicle control

Westlake University and collaborators presented a dual-brain vision-language-action architecture for autonomous underwater navigation. A multimodal model handles high-level reasoning while hydrodynamics-informed model-predictive control handles fast low-level response. Authors report 19–27% higher task completion than baselines in field tests under degraded visibility.

Why it matters: Separating probabilistic reasoning from deterministic/reactive control is a practical safety pattern for embodied AI with limited communications.

Impact: Results need independent replication across vehicles, currents, sensors and mission types before operational reliance. Scientific Reports

Healthcare AI · Singapore/China