TANGO: Humanoid Navigation in Cluttered Environments with a Whole-Body Vision-Language-Action Model
TANGO is introduced as the first whole-body vision-language navigation framework for language-conditioned humanoid traversal in cluttered environments. It directly predicts 29-DoF joint-space actions from natural-language instructions and egocentric RGB observations. The model is trained entirely in simulation using a pipeline that synthesizes collision-free traversal behaviors via global path planning, kinematic whole-body motion generation, obstacle-aware motion editing, and RL-based tracking. In simulation experiments, TANGO demonstrates state-of-the-art performance in vision-language navigation and outperforms strong modular baselines.
Researchers present TANGO, a whole-body vision-language-action model for humanoid robots navigating cluttered indoor spaces. Unlike conventional 2D path planning, TANGO performs continuous geometry-aware whole-body adaptation, coordinating arm placement, torso adjustment, and gait modulation. Given a language instruction and egocentric RGB input, the model outputs 29-DoF joint-space actions. Training is simulation-only, using a pipeline that generates diverse collision-free trajectories and provides dynamically feasible action supervision. Simulation results show state-of-the-art vision-language navigation performance, surpassing modular baselines.
TANGO integrates vision-language understanding with whole-body control by directly mapping language and RGB observations to 29-DoF joint-space actions. The simulation training pipeline combines global path planning, kinematic motion generation, obstacle-aware editing, and reinforcement learning-based tracking to produce feasible supervision. This end-to-end approach avoids explicit modular decomposition and enables coordinated whole-body adaptation in complex 3D environments.
The work signals progress toward deployable humanoid robots capable of following natural-language commands in real-world cluttered settings. Simulation-only training reduces the need for costly real-world data collection, potentially accelerating development cycles. However, the gap between simulation and physical deployment remains a key challenge, and the absence of real-world validation limits immediate commercial readiness.
If successfully transferred to physical robots, TANGO could enable humanoid robots for warehouse navigation, service robotics, and assistive tasks in cluttered indoor environments. The simulation-based training approach may lower development costs and time-to-market for robot navigation capabilities.
Next observable signals include real-world demonstrations of TANGO on physical humanoid platforms, comparisons with additional baselines, and extensions to more diverse environments and instructions. Further research may address sim-to-real transfer, safety constraints, and integration with higher-level task planning.