OWLS: Open-world Language-conditioned Speed Fields for Zero-Shot Rover Navigation

TL;DR

OWLS enables zero-shot rover navigation by using open-vocabulary vision-language models to generate speed fields directly from camera images, planning trajectories based on terrain appearance without task-specific training or annotated data. It adapts VLM outputs via morphological operations to guide minimum-time planners across diverse environments.

Why It Matters

This approach eliminates the need for labeled terrain data, persistent maps, or additional sensors, making it highly practical for planetary exploration where training data and infrastructure are limited. It enables adaptable, real-time navigation in open-world environments using only a single camera.

Key Findings

• OWLS achieves zero-shot navigation without any task-specific training or terrain annotations.
• The method generalizes across diverse terrains, including simulated Mars and Moon environments and real planetary imagery.
• Morphological operations effectively adapt VLM scores into robust speed fields for safe, minimum-time trajectory planning.