AI-Driven Dynamic Covariance for ROS 2 Mobile Robot Localization

Abstract

In the evolving field of mobile robotics, enhancing localization robustness in dynamic environments remains a critical challenge, particularly for ROS 2-based systems where sensor fusion plays a pivotal role. This study evaluates an AI-driven approach to dynamically adjust covariance parameters for improved pose estimation in a differential-drive mobile robot. A regression model was integrated into the robot_localization package to adapt the Extended Kalman Filter (EKF) covariance in real time, with experiments conducted in a controlled indoor setting over runs comparing AI-enabled dynamic covariance prediction against a static covariance baseline across Static, Moderate, and Aggressive motion dynamics. The AI-enabled system achieved a Mean Absolute Error (MAE) of 0.0061 for pose estimation and reduced median yaw prediction errors to 0.0362 rad (static) and 0.0381 rad (moderate) with tighter interquartile ranges (0.0489 rad, 0.1069 rad) compared to the baseline (0.0222 rad, 0.1399 rad). Aggressive dynamics posed challenges, with errors up to 0.9491 rad due to data distribution bias and Random Forest model constraints. Enhanced dataset augmentation, LSTM modeling, and online learning are proposed to address these limitations. Datalogging enabled iterative re-training, supporting scalable state estimation with future focus on online learning.