Performance Evaluation of Machine Learning Models for Mobile Robot Odometry Error Prediction Using Gazebo

Precise localisation is a basic requirement for autonomous mobile robots operating in complex indoor environments. The wheel encoder based odometry is the most widely used dead-reckoning technique because of its simplicity, cheap implementation cost and great computing efficiency. However, the readings from the wheel encoder are very sensitive to the accumulated errors due to physical wheel slippage, structural misalignment, uneven surface terrains and deterministic mechanical flaws. Such unchecked concerns will cause localisation drift and may degrade the stability of the route planning and obstacle avoidance behaviours. This paper systematically evaluates the performance of three state-of-the-art gradient boosting frameworks, namely XGBoost, LightGBM and CatBoost, for the dynamic prediction and correction of cumulative localisation errors.