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

Anqi Li1,2,* Yuxin Chen1,*,‡ Zhaobo Li1,* Zhuo Cao1,3,* Junli Ren1,4
Masayoshi Tomizuka1 Dhruv Shah5,†
1 University of California, Berkeley 2 Peking University 3 Tsinghua University
4 The University of Hong Kong 5 Princeton University
* Equal contribution. ‡ Project lead. † Corresponding author.
Paper Video
RSS@WCBM Oral

TANGO is a whole-body vision-language-action framework for language-conditioned humanoid traversal in cluttered indoor environments.

Abstract

We study the problem of navigating cluttered indoor environments with a humanoid robot. Unlike conventional navigation methods that operate over simplified navigation abstractions, humanoid traversal in cluttered environments requires continuous geometry-aware whole-body adaptation, including coordinated arm placement, torso adjustment, and gait modulation for collision-free movement through complex 3D spaces.

We introduce TANGO, a whole-body vision-language navigation framework for language-conditioned humanoid traversal in cluttered environments. Given a natural-language instruction and egocentric RGB observations, TANGO predicts 29-DoF joint-space actions for whole-body humanoid control.

TANGO is trained entirely in simulation by synthesizing diverse collision-free traversal behaviors through global path planning, kinematic whole-body motion generation, obstacle-aware motion editing, and reinforcement-learning-based tracking. In simulation and real-world experiments, TANGO improves language-guided navigation and transfers zero-shot to a Unitree G1 humanoid without real-world training data.

Method

Traversal via Whole-Body Vision-Language-Action Model

TANGO is a whole-body VLA system for cluttered indoor vision-language navigation. It combines a simulation data pipeline for collision-free traversal supervision with a triple-system model that maps language and egocentric RGB observations to executable whole-body action chunks.

TANGO architecture and data-generation pipeline.
TANGO architecture and data pipeline. Plan-Edit-Track synthesizes whole-body traversal data; a vision-language backbone, action expert, and low-level tracker produce continuous humanoid control.

Problem Formulation

Given a language instruction, front/downward RGB history, and whole-body proprioception, TANGO predicts a horizon of whole-body actions: 29-DoF desired joint angles plus a 6D base rotation representation.

Simulation Data Generation

VLNVerse and SAGE-3D scenes are augmented with lateral, ground-level, and overhead obstacles. The PET pipeline plans paths, edits motions for obstacle avoidance, and tracks them to dynamic feasibility.

TANGO Architecture and Training

A 7B vision-language backbone is paired with a flow-matching MM-DiT action expert. Real-time chunking aligns training with streaming execution through a robust low-level motion tracker.

Deployment

At deployment, the low-frequency VLA module runs on a server while the high-frequency whole-body tracker runs onboard, enabling responsive real-world humanoid control.

TANGO real-world deployment system.
Real-world deployment system. A cloud-edge setup separates low-frequency VLA inference from high-frequency whole-body tracking.

Experiments

Simulation and Real-World Evaluation

VLN Performance

Ours is the only compared method evaluated with low-level physical control, yet achieves the best SR and NE on both seen and unseen VLNVerse validation splits.

Cluttered Environment Traversal Performance

On augmented cluttered VLNVerse scenes, Ours improves SR and SPL while reducing collision rate to 9.90 percent using only RGB observations.

Real-World Experiment

Ours transfers zero-shot to a Unitree G1 humanoid for long-horizon navigation, cluttered-scene traversal, and narrow-passage side-stepping.

Ablation Studies

Whole-body action generation remains robust under physical low-level control, while planar policies degrade substantially when moved from teleportation to embodied execution.

VLNVerse Benchmark

Method Low-Level Control Val Seen Val Unseen
NE OSR SR SPL NE OSR SR SPL
CMA No 5.36 59.81 37.35 33.36 5.16 62.79 31.15 27.92
RDP No 4.02 68.09 47.28 41.69 3.75 71.93 48.60 42.72
Seq2Seq No 4.78 44.68 32.62 30.39 4.36 49.58 35.03 33.37
Ours Yes 3.90 70.31 54.69 40.18 3.72 71.07 52.89 40.18

Cluttered VLNVerse Scenes

Method NE SR SPL CR
InternVLA-N1 + Unitree Control 5.34 26.67 11.92 19.03
InternVLA-N1 + HumanoidPF 3.19 35.29 24.35 16.11
Ours 4.01 43.75 31.83 9.90

Action-Space Ablation

Method Low-Level Control SR SPL
InternVLA-N1 No 43.69 35.74
Ours-2D No 45.74 35.44
InternVLA-N1 Yes 42.37 22.50
Ours-2D Yes 26.67 8.34
Ours Yes 52.89 40.18
Real-world humanoid navigation results.
Real-world deployment. Ours performs long-horizon language-guided navigation, traverses cluttered environments, and adapts whole-body motions to pass through narrow spaces without real-world training.

Conclusion

Ours demonstrates that a whole-body VLA policy can directly predict 29-DoF actions for humanoid navigation. The results show strong language-guided navigation performance, improved collision-aware traversal in cluttered scenes, and zero-shot sim-to-real transfer on humanoid hardware.