Research | Aaron Wilhelm | Cornell University

Research Overview

High-precision localization is a fundamental ability required to enable higher-level autonomous behavior in robots. While accurate localization can be achieved in relatively controlled environments, performance in GPS-denied, dynamic settings remains a significant challenge. Traditional localization methods using LiDAR or outward-facing cameras, for example, can struggle in these scenarios due to occlusions from moving objects and people, as well as variations in ambient lighting.

My PhD research focuses on an alternative and highly promising modality: ground texture localization. This approach utilizes a downward-facing camera to localize a robot based on the unique visual appearance of the ground surface. Because the camera’s view is directed downward, it is largely unaffected by outward obstructions and can be paired with a controlled lighting system, mitigating many common environmental challenges.

This methodology offers several distinct advantages:

High precision: Achieves submillimeter-level localization accuracy.
Robustness to dynamic environments: Less affected by moving objects or people.
No environmental modification: Requires no markers, beacons, or infrastructure changes.
Complementary to existing methods: Can be integrated with other localization techniques.

Through my research, I aim to advance the field of robot localization and mapping, enabling more capable and reliable autonomous systems that can operate effectively in dynamic, real-world environments.

Research Projects

Geometrically-Constrained Ground Texture SLAM Pipeline

My current work is exploring how geometric constraints can be be exploited throughout the entire SLAM pipeline to enhance localization accuracy and robustness. This work is currently in preparation for publication.

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

I developed a significantly improved bag-of-words (BoW) image retrieval system specifically designed for ground texture localization. This system achieves substantially higher accuracy for global localization and higher precision and recall for loop closure detection in SLAM.

Key Contributions:

Leverages an approximate k-means (AKM) vocabulary with soft assignment
Exploits consistent orientation and constant scale constraints inherent to ground texture localization
Provides both high-accuracy and high-speed versions for different use cases

Performance Highlights:

Improved mean average precision from 0.026 to 0.559 for global localization
Identified almost 3x as many loop closures for SLAM applications

Lightweight Ground Texture Localization (L-GROUT)

I developed a lightweight ground texture localization algorithm that improves the state of the art in both performance and computational efficiency. This algorithm can run in real-time on single board computers without GPU acceleration, making it ideal for small indoor robots.

Key Contributions:

Improved database feature extraction algorithm
Dimensionality reduction method based on locality preserving projections (LPP)
Enhanced spatial filtering for greater robustness

Performance Highlights:

Achieved localization frequencies of ~4Hz on a Raspberry Pi 4

This work enables high-precision, millimeter-level localization without instrumenting, marking, or modifying the environment, making it accessible for a wide range of robotic applications.

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

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 and future research aims to further advance ground texture localization and SLAM techniques:

Improving Localization Precision: Systematically exploring novel approaches to push localization precision beyond the millimeter level, targeting micrometer-scale accuracy.
Deep Learning Integration: Integrating deep learning approaches for improved feature extraction and matching.
Comprehensive Datasets: Creating comprehensive datasets to facilitate research in this area.