OWLS: Open-world Language-conditioned Speed Fields for Zero-Shot Rover Navigation
OWLS: Open-world Language-conditioned Speed Fields for Zero-Shot Rover Navigation
OWLS is a novel approach that uses open-vocabulary vision-language models to generate speed fields for rover navigation. By interpreting descriptions of traversable terrain, OWLS scores each pixel of a camera image, projects these scores onto the ground, and adapts them for the platform to generate a minimum-time trajectory. It operates without task-specific training, terrain annotation, persistent maps, or exteroceptive sensors beyond a single camera, and is evaluated on both physical and simulated rovers across diverse terrains.
OWLS leverages open-vocabulary vision-language models to enable zero-shot navigation for rovers by translating terrain descriptions into speed fields, eliminating the need for task-specific training or sensor data beyond a single camera.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
Source：arXiv Robotics — research abstracts · arxiv.org