OnFly-Outdoor is a zero-shot navigation framework that enables robots to follow natural language instructions in unstructured outdoor environments while avoiding moving obstacles. It integrates multimodal perception, dynamic motion forecasting, and language-grounded reactive planning to address outdoor navigation challenges.
Key findings
OnFly-Outdoor achieves a 35% improvement in success rate over baselines in dynamic scenarios.
The framework maintains real-time inference at 10Hz.
The work establishes a foundation for robust outdoor AVLN and identifies critical research directions for physical robot deployment.
Limitations & open questions
The framework's performance in real-world deployment is yet to be tested.
The current implementation may require further optimization for physical robot integration.