Core idea
Combines classical frontier exploration with egocentric VLM reasoning and a global 3D semantic map (OpenRoboVox); the bandit shifts the robot from exploration to semantically-guided navigation as understanding improves. 90.6% on GOAT-Bench Val Unseen (+17.8pp), runs on a real Unitree Go2 in 1800m² spaces — suggesting semantic grounding can matter more than model size.