TANGO: Humanoid Navigation in Cluttered Environments with a Whole-Body Vision-Language-Action Model
Anqi Li, Yuxin Chen, Zhaobo Li, Zhuo Cao, Junli Ren, Masayoshi Tomizuka, Dhruv Shah
Read on arXiv →
Key claim
TANGO enables language-guided humanoid navigation in real environments.
In plain English
Imagine you're building a humanoid robot that needs to navigate through a crowded room filled with furniture and obstacles. Traditional navigation methods often rely on 2D path planning, which can lead to issues when the robot encounters complex 3D environments. For instance, a robot might get stuck or collide with objects because it doesn't adapt its movements to the three-dimensional space around it. This is what's called a failure in whole-body adaptation, where the robot struggles to coordinate its arms, torso, and legs effectively. To address this, TANGO offers a fresh approach by combining vision and language to guide the robot's movements. Instead of just following a flat path, TANGO allows the robot to understand natural language instructions and make real-time adjustments based on what it sees. It predicts joint-space actions for its entire body, enabling it to navigate through cluttered environments smoothly. The training process occurs entirely in simulation, where diverse movement behaviors are synthesized, ensuring the robot learns to avoid collisions and adapt its actions dynamically. Compared to previous methods, TANGO not only shows superior performance in simulated environments but also successfully operates in real-world scenarios without needing prior training on actual navigation data. This means that for builders, TANGO represents a significant step forward in creating robots that can understand and navigate complex spaces using human-like instructions.
TANGO introduces a novel approach to humanoid navigation by integrating vision and language for whole-body control.
The evaluation against strong baselines and zero-shot deployment on a real robot supports the findings.
Deep reliability assessment
The methodology supports the claim of robust language-guided traversal in cluttered environments through simulation-based training and zero-shot deployment on real hardware. However, the lack of real-world training data might limit the generalizability of the results.
Reproducibility
yes, the paper mentions a project website: tango-vla.github.io
Key figure
Figure 1 illustrates humanoid whole-body navigation in cluttered environments, showcasing TANGO's capability to navigate a 30-meter route zero-shot in a real-world cluttered office scene.