Learning Humanoid Navigation from Human Data
Summary: arXiv:2604.00416v1 Announce Type: cross
Abstract
We present EgoNav, a groundbreaking system designed to enable humanoid robots to navigate through diverse, unseen environments by learning solely from 5 hours of human walking data. This innovative approach does not require any robot-specific data or finetuning.
Key Features of EgoNav
- Diffusion Model: A diffusion model is employed to predict distributions of plausible future trajectories conditioned on past trajectories.
- 360-Degree Visual Memory: The system integrates a 360-degree visual memory that fuses color, depth, and semantic information.
- Advanced Video Features: It utilizes video features from a frozen DINOv3 backbone, which captures appearance cues that are typically invisible to traditional depth sensors.
- Hybrid Sampling Scheme: A hybrid sampling scheme allows for real-time inference in just 10 denoising steps.
- Receding-Horizon Controller: A receding-horizon controller effectively selects paths from the predicted distribution.
Validation and Performance
The performance of EgoNav has been rigorously validated through offline evaluations, where it consistently outperformed existing baselines in critical areas such as collision avoidance and multi-modal coverage. Additionally, the system has been tested through zero-shot deployment on a Unitree G1 humanoid robot across various unseen indoor and outdoor environments.
Emergent Behaviors
As a result of the learning process, EgoNav exhibits a variety of sophisticated behaviors that emerge naturally from its learned prior. These behaviors include:
- Waiting for doors to open before proceeding.
- Navigating around crowds with awareness of human presence.
- Avoiding obstacles such as glass walls, demonstrating advanced spatial awareness.
Future Directions
In our commitment to advancing the field of humanoid navigation, we will be releasing the dataset and trained models associated with EgoNav. This step will enable researchers and developers to build upon our findings, fostering innovation in robotic navigation systems.
Visit Us
For more information about EgoNav, please visit our website: https://egonav.weizhuowang.com