A Comparison of Imitation Learning Pipelines for Autonomous Driving on the Effect of Change in Ego-vehicle

This paper presents a comparison of the effect of change in ego-vehicle in two different pipelines of imitation learning for autonomous driving (AD) between direct control-based and waypoint-based pipelines. Control-based pipeline involves predicting control signals directly to control the car, whereas a waypoint-based pipeline predicts the future trajectory of the car and uses a controller module to generate the control signals from the predicted waypoints. In this study, CIL++ was used for the control-based method whereas TransFuser was used for the waypoint-based method. In our experiments, we used CARLA simulator and deployed both imitation learning models, without retraining or re-tuning the controller parameters, on various cars different from the car used during training. We used Town05 from CARLA’s Leaderboard benchmark to evaluate the performance based on driving score, the main metric used in the benchmark. Based on the experiment results, TransFuser is more robust in adapting to different ego-vehicles than CIL++. TransFuser performed better when deployed to different vehicles. However, the performance still suffered when there was a significant change in the car classes. The source code of this work is made publicly available at https://github.com/ailabspace/Comparisonof-Autonomous-Driving-IL-Pipeline-for-Ego-Vehicle-Changes.