Research | Aaron Wilhelm | Cornell University

Research Overview

Reliable localization is a fundamental capability for autonomous robots, yet achieving high precision in GPS-denied, dynamic, and visually challenging environments remains difficult. My research develops accurate, robust, and computationally efficient localization and mapping methods for mobile robots.

My Ph.D. work centers on ground-texture localization and SLAM, which use a downward-facing camera to localize a robot from the natural visual texture of the surface beneath it. By exploiting the constrained geometry and controlled sensing conditions of this setup, I have developed lightweight localization and visual place-recognition methods that achieve submillimeter precision without modifying the environment while remaining robust to occlusions and dynamic surroundings.

My current work extends these ideas toward scalable hybrid SLAM across surfaces with varying texture and microscope-based localization with accuracy on the order of tens of micrometers.

Research Projects

Most projects include a hands-on interactive demo you can run right in your browser — open a project to try it.

Scalable Ground-Texture SLAM

My current work develops a hybrid loop-closure architecture for scalable ground-texture SLAM. The system combines complementary localization methods to maintain robust tracking across surfaces with varying texture while limiting the memory required for each keyframe.

In experiments to date, this approach reduces per-keyframe memory by up to 55× relative to a direct-method baseline. A manuscript describing the system is in preparation.

Interactive demo

Improved Bag-of-Words Image Retrieval for Ground Texture Localization

GT-BoW is a geometry-aware bag-of-words image-retrieval system for ground-texture localization. It uses the fixed scale and consistent orientation of downward-facing imagery to reject visually similar but geometrically inconsistent matches, improving both global localization and loop-closure detection.

Key Contributions

An approximate k-means vocabulary with soft assignment
Scale-binned vocabularies and orientation-based geometric verification
Separate high-accuracy and high-speed variants for global localization and loop closure

Performance Highlights

Increased global-localization mean average precision from 0.026 to 0.559
Detected nearly 3× as many loop closures as a standard DBoW baseline

Explore the live demo

Interactive demo

Lightweight Ground Texture Localization (L-GROUT)

L-GROUT is a lightweight global-localization system that matches natural ground texture observed by a downward-facing camera to a prebuilt map. It improves localization performance while reducing computation enough to run in real time on resource-constrained, single-board computers without GPU acceleration.

Key Contributions

An improved method for selecting database features
Locality-preserving dimensionality reduction that supports fast binary features
A spatial-voting stage designed to retain accuracy in lightweight configurations

Performance Highlight

Runs the complete localization pipeline at more than 4 Hz on a Raspberry Pi 4

L-GROUT provides millimeter-scale global localization without markers, installed infrastructure, or other environmental modifications.

Related Publication: "Lightweight Ground Texture Localization" (ICRA 2024)

Explore the live demo

Applications

Warehouse Robotics

Precise navigation in dynamic environments with frequent human interaction and moving objects.

Autonomous Vehicles

Complementary localization in areas where GPS is unreliable or unavailable.

Indoor Service Robots

Reliable positioning without modifying the environment for service and delivery robots.

Future Research Directions

My ongoing research extends ground-texture localization toward new sensing scales, more varied surfaces, and stronger simulation tools:

Micrometer-scale localization: Using microscopic surface texture and low-cost hardware to localize mobile robots with accuracy on the order of tens of micrometers.
Scalable hybrid SLAM: Maintaining reliable tracking across low-texture regions while reducing the memory required for loop closure.
Ground-texture simulation: Generating controllable, spatially consistent imagery for localization and SLAM experiments.