TANGO: Humanoid Navigation in Cluttered Environments with a Whole-Body Vision-Language-Action Model

TANGO. Humanoid Navigation in Cluttered Environments with a Whole-Body Vision-Language-Action Model. By Anqi Li and six others, from the University of California, Berkeley, Peking University, and three other institutions. Published on the eighth of September. No code repository is linked to it.

The problem is a humanoid robot navigating a cluttered indoor environment. Conventional methods treat navigation as path planning in 2D. A humanoid moving through clutter needs continuous whole-body adaptation that is aware of geometry: placing its arms in coordination, adjusting its torso, and modulating its gait, so it can move through complex 3D spaces without collision. A route drawn on a floor plan says where to go, but not how a body with arms, a torso and legs should fit through the gaps along the way.

TANGO is presented as the first whole-body vision-language navigation framework for humanoid traversal of cluttered environments, conditioned on language. Given a natural-language instruction and egocentric RGB observations, it directly predicts actions in joint space, twenty-nine degrees of freedom, for downstream whole-body control. There is no separate 2D plan to follow; the model outputs how the body moves. Language says where to go, the camera shows what is in the way, and the output is the posture and motion of the whole body.

TANGO is trained entirely in simulation. Diverse collision-free traversal behaviours are synthesised through a sequence of steps: global path planning, kinematic generation of whole-body motion, motion editing that accounts for obstacles, and tracking based on reinforcement learning. That pipeline provides action supervision that is dynamically feasible, which is what a language-conditioned whole-body policy needs to learn from. Each stage adds something the last one lacked: a route, then a body moving along it, then a motion adjusted to the obstacles, then tracking of that motion learned through reinforcement.

In extensive simulation experiments, TANGO reaches state-of-the-art performance in vision-language navigation. It outperforms strong modular baselines in challenging scenes that require negotiating obstacles. Finally, it is deployed zero-shot on a Unitree G1 humanoid robot, and shows robust language-guided traversal through cluttered real-world scenes, with no training on any real-world navigation data.

The real-robot result closes the argument. A policy learned only from synthesised simulation behaviour, predicting joints directly from what the robot sees and is told, carries over to a physical humanoid in real clutter. The supervision pipeline is what made that possible without collecting real navigation data. And against modular systems that split the problem into parts, the single whole-body model does better precisely in the scenes where obstacles have to be negotiated.