Is this project an undergraduate, graduate, or faculty project?
Undergraduate
Project Type
group
Campus
Daytona Beach
Authors' Class Standing
Jose Demedeiros, Senior
Lead Presenter's Name
Jose Demedeiros
Lead Presenter's College
DB College of Engineering
Faculty Mentor Name
Dr. Cagri Kilic
Abstract
This work proposes an energy-aware meta-learning framework that selects the single most suitable localization algorithm for a lunar rover, per scene, using only monocular imagery and orbital maps. The goal is to achieve sub-meter accuracy while minimizing onboard compute and energy consumption. We assemble a suite of seven lunar-relevant algorithms spanning relative and absolute localization, including monocular ORB-SLAM3, LuVo homography-based visual odometry, Censible cross-view matching with orbital imagery, crater-based methods (LunarNav and ShadowNav), monocular horizon navigation with a DEM, and DROID-SLAM. Relative methods provide incremental motion updates, while absolute methods deliver global pose fixes; an Extended Kalman Filter fuses these without relying on IMU propagation, avoiding double-integration noise. A three-branch neural meta-selector chooses which single algorithm to run at each frame. Its inputs are: (i) a scene context vector capturing feature density, texture entropy, shadow fraction, slope proxy, and orbital match availability; (ii) a Vision Transformer (ViT-Tiny) embedding of the current frame; and (iii) a per-algorithm energy-cost vector derived from runtime and power profiling. In an offline calibration phase with ground truth, each algorithm’s performance is summarized by an accuracy–efficiency score combining Absolute Trajectory Error with normalized energy cost. The algorithm with the highest score in each scene serves as the training label, and the selector is trained using cross-entropy. At deployment, the selector runs in real time, and is adapted online using an incremental meta-learning procedure driven by consistency between occasional paired algorithm runs, removing any need for ground truth during operations. Simulation experiments in a photorealistic lunar environment evaluate per-algorithm performance across crater-rich, flat regolith, rock-dense, shadowed, and transitional scenes, and quantify both selector regret relative to an oracle policy and energy savings compared to always running all algorithms in parallel.
Did this research project receive funding support (Spark, SURF, Research Abroad, Student Internal Grants, Collaborative, Climbing, or Ignite Grants) from the Office of Undergraduate Research?
No
Included in
Artificial Intelligence and Robotics Commons, Space Vehicles Commons, Systems Engineering and Multidisciplinary Design Optimization Commons
An Energy-Aware Meta-Learning Framework for Real-Time Lunar Rover Localization via Adaptive Algorithm Selection
This work proposes an energy-aware meta-learning framework that selects the single most suitable localization algorithm for a lunar rover, per scene, using only monocular imagery and orbital maps. The goal is to achieve sub-meter accuracy while minimizing onboard compute and energy consumption. We assemble a suite of seven lunar-relevant algorithms spanning relative and absolute localization, including monocular ORB-SLAM3, LuVo homography-based visual odometry, Censible cross-view matching with orbital imagery, crater-based methods (LunarNav and ShadowNav), monocular horizon navigation with a DEM, and DROID-SLAM. Relative methods provide incremental motion updates, while absolute methods deliver global pose fixes; an Extended Kalman Filter fuses these without relying on IMU propagation, avoiding double-integration noise. A three-branch neural meta-selector chooses which single algorithm to run at each frame. Its inputs are: (i) a scene context vector capturing feature density, texture entropy, shadow fraction, slope proxy, and orbital match availability; (ii) a Vision Transformer (ViT-Tiny) embedding of the current frame; and (iii) a per-algorithm energy-cost vector derived from runtime and power profiling. In an offline calibration phase with ground truth, each algorithm’s performance is summarized by an accuracy–efficiency score combining Absolute Trajectory Error with normalized energy cost. The algorithm with the highest score in each scene serves as the training label, and the selector is trained using cross-entropy. At deployment, the selector runs in real time, and is adapted online using an incremental meta-learning procedure driven by consistency between occasional paired algorithm runs, removing any need for ground truth during operations. Simulation experiments in a photorealistic lunar environment evaluate per-algorithm performance across crater-rich, flat regolith, rock-dense, shadowed, and transitional scenes, and quantify both selector regret relative to an oracle policy and energy savings compared to always running all algorithms in parallel.