Title:    A Multi-Rate, Role-Based Architecture for Visual SLAM
Date:   Friday, October 30th, 2026
Time:  12:00PM - 1:30PM EDT
Location:  TSRB 523a
Virtual Link: Teams

Yanwei Du
Robotics PhD Student
School of Electrical and Computer Engineering
Georgia Institute of Technology

Committee members
•    Dr. Patricio A. Vela (advisor), School of Electrical and Computer Engineering, Georgia Institute of Technology
•    Dr. Matthew Hale,                       School of Electrical and Computer Engineering, Georgia Institute of Technology
•    Dr. Anthony Yezzi,                       School of Electrical and Computer Engineering, Georgia Institute of Technology
•    Dr. Lu Gan,                                     School of Aerospace Engineering, Georgia Institute of Technology
•    Dr. Frank Dellaert,                      School of Interactive Computing, Georgia Institute of Technology

Abstract:

Autonomous navigation requires accurate and timely pose estimates throughout task execution. This thesis develops a multi-rate, role-based SLAM architecture in which inertial prediction, direct tracking, and indirect feature-based estimation operate at different rates. It investigates whether feedforward motion priors and feedback corrections between these estimators can improve tracking continuity and reduce drift.

In the feedforward path, inertial prediction provides a motion prior to direct tracking, which in turn supplies a prior to indirect estimation. In the feedback path, indirect estimation returns corrected keyframe poses to direct tracking and updates the inertial state through a separate preintegration-based correction. Adaptive weights determine the influence of each prior and correction. Feedforward weights account for the quality of both the receiving estimator and the incoming prior, while feedback weights account for keyframe covisibility and how strongly the direct tracker is already constrained. Earlier results from Good Weights and Good Baseline provide evidence that adaptive motion priors and asynchronous refinement of keyframe poses can improve feature-based SLAM, motivating their integration and extension within this architecture.

The primary contributions are the unified architecture and its adaptive weighting strategies. Open-loop experiments evaluate trajectory accuracy, tracking continuity, and latency across computing devices, while closed-loop navigation experiments assess improvements in waypoint completeness and precision.