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

What it claims

TANGO is presented as the first whole-body vision-language navigation framework for language-conditioned humanoid traversal in cluttered indoor spaces. Its central argument against prior work: humanoid navigation is not a 2D path-planning problem but requires continuous geometry-aware whole-body adaptation — coordinated arm placement, torso adjustment, and gait modulation to move collision-free through 3D clutter. Given a natural-language instruction plus egocentric RGB, TANGO directly predicts 29-DoF joint-space actions for downstream whole-body control, rather than emitting a 2D path or waypoints for a separate controller.

How it works

The model is trained entirely in simulation. The authors synthesize diverse collision-free traversal behaviors through a four-stage pipeline that produces dynamically feasible action supervision for the language-conditioned policy:

flowchart LR
  A[Global path planning] --> B[Kinematic whole-body motion generation]
  B --> C[Obstacle-aware motion editing]
  C --> D[RL-based tracking]
  D --> E[Feasible 29-DoF action supervision]
  E --> F[Language-conditioned whole-body policy]

Global path planning fixes the route; kinematic generation produces whole-body motion along it; obstacle-aware editing adjusts that motion for clutter; RL tracking makes the edited motion dynamically feasible. The resulting trajectories become supervision for the VLA policy. Inputs are the instruction and egocentric RGB; there is no stated dependence on a separate map or depth sensor in the abstract.

Evidence

The page reports only qualitative claims, no numbers. In simulation, TANGO is stated to reach state-of-the-art vision-language navigation performance and to outperform strong modular baselines on scenes requiring obstacle negotiation. It was then deployed zero-shot on a Unitree G1 humanoid, with robust language-guided traversal in cluttered real-world scenes and no training on any real-world navigation data — the sim-to-real transfer is the headline empirical claim.

Limits

This is an abstract only; no benchmark names, metric values, baseline identities, dataset sizes, compute, or code/weights are given. The real-world result is described qualitatively ("robust") without success rates or trial counts, and only one platform (Unitree G1) is named. Whether the modular baselines were tuned comparably, and what the actual SOTA margins are, cannot be judged from this page — an engineer would need the PDF for any adoption decision.