Uncertainty Quantification and Calibration of Imitation Learning Policy in Autonomous Driving

Abstract

Current state-of-the-art imitation learning policies in autonomous driving, despite having good driving performance, do not consider the uncertainty in their predicted action. Using such an unleashed action without considering the degree of confidence in a blackbox machine learning system can compromise safety and reliability in safety-critical applications such as autonomous driving. In this paper, we propose three different uncertainty-aware policies, to capture epistemic and aleatoric uncertainty over the continuous control commands. More specifically, we extend a state-of-the-art policy with three common uncertainty estimation methods: heteroscedastic aleatoric, MC-Dropout and Deep Ensembles. To provide accurate and calibrated uncertainty, we further combine our agents with isotonic regression, an existing calibration method in regression task. We benchmark and compare the driving performance of our uncertainty-aware agents in complex urban driving environments. Moreover, we evaluate the quality of predicted uncertainty before and after recalibration. The experimental results show that our Ensemble agent combined with isotonic regression not only provides accurate uncertainty for its predictions but also significantly outperforms the state-of-the-art baseline in driving performance.