Humanoid Robots Navigate Narrow Gaps and Obstacles With Whole-Body AI Control
TANGO is a whole-body vision-language-action model from UC Berkeley and Princeton that takes a spoken instruction and a camera feed and outputs motion for all 29 joints of a humanoid, rather than planning a flat two-dimensional path. Trained on roughly 65,000 physics-verified trajectories synthesised in about 211 GPU-hours, it succeeded 53% of the time in simulation against 27% for the 2D baseline, and cut real-robot collisions from about 16% to about 10% using RGB cameras alone. This article covers the architecture, the arithmetic and the stated limits.
