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.
TANGO
Humanoid Navigation in Cluttered Environments
with a Whole-Body Vision-Language-Action Model
Conference on Robot Learning 2026
Navigation is a whole-body problem.
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.
From language to whole-body motion.
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.
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.
One humanoid. A whole body of possibilities.
Long-horizon navigation and geometry-aware traversal on a Unitree G1, transferred zero-shot from simulation.
Making room with whole-body motion
Sideways, under, and over: adapting to the geometry of a passage.
Sidestep
Reorienting the body to pass through a narrow opening.
Squat
Lowering the body to move beneath an overhead obstacle.
Stride
Stepping over a ground-level obstacle.
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.
BibTeX
@inproceedings{li2026tango,
title = {{TANGO}: Humanoid Navigation in Cluttered Environments
with a Whole-Body Vision-Language-Action Model},
author = {Li, Anqi and Chen, Yuxin and Li, Zhaobo and Cao, Zhuo and
Ren, Junli and Tomizuka, Masayoshi and Shah, Dhruv},
booktitle = {Conference on Robot Learning (CoRL)},
year = {2026},
url = {https://arxiv.org/abs/2609.09158}
}