Zero-shot semantic landmark-based visual odometry using foundation models for unstructured planetary exploration

Area of Science:

• Robotics
• Computer Vision
• Planetary Science

Background:

• Precise autonomous navigation is crucial for space exploration in GNSS-denied areas.
• Traditional Visual Odometry (VO) struggles with extreme lighting and monotonous terrains found on planetary surfaces.
• Existing methods often fail due to reliance on low-level geometric features.

Purpose of the Study:

• To develop a zero-shot semantic landmark-based VO approach for planetary navigation.
• To leverage Foundation Models for robust feature extraction and matching in extraterrestrial environments.
• To evaluate the performance of the proposed method in diverse, challenging planetary analog conditions.

Main Methods:

• Utilized the Segment Anything Model (SAM) for geological landmark (rock) extraction.
• Employed DINOv2 for generating view-invariant semantic descriptors.
• Matched semantic landmarks across image frames for pose estimation.
• Evaluated on synthetic lunar (LuSNAR) and real-world Martian analog (Katwijk Beach) datasets.

Main Results:

• Achieved decimeter-level trajectory accuracy (RMSE ≈ 0.14 m) on a Martian analog dataset.
• Demonstrated an RMSE of 1.93 m on a stable lunar traverse.
• Showcased robustness against extreme illumination and textural monotony without domain-specific fine-tuning.
• Validated zero-shot transfer capabilities across different planetary domains.

Conclusions:

• Foundation Model-based semantic landmarks offer a promising alternative to traditional low-level features for VO.
• The proposed approach enables precise autonomous navigation in challenging planetary environments.
• This method advances the potential for future lunar and Martian exploration missions.