Prediction-Uncertainty-Aware Decision-Making for Autonomous Vehicles

Abstract

Motion prediction is the fundamental input for decision-making in autonomous vehicles. The current motion prediction solutions are designed with a strong reliance on black box predictions based on neural networks (NNs), which is unacceptable for safety-critical applications. Motion prediction with high uncertainty can cause conflicting decisions and even catastrophic results. To address this issue, an uncertainty estimation approach based on the deep ensemble technique is proposed for motion prediction in this paper. Subsequently, the estimated uncertainty is considered in the decision-making module to improve driving safety. Firstly, a motion prediction model based on long short-term memory (LSTM) is built and the deep ensemble technique is utilized to obtain both epistemic and aleatoric uncertainty of the prediction model. Besides, an uncertainty-aware potential field is developed to process the prediction uncertainty. Furthermore, a decision-making framework is proposed based on the model predictive control algorithm that considers the uncertainty-aware potential field, road boundaries, and multiple constraints of vehicle dynamics. Finally, the public available <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">NGSIM</i> , <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">HighD</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">INTERACTION</i> datasets are used to evaluate the proposed motion prediction model. More importantly, two traffic scenarios are also extracted from <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">NGSIM</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">INTERACTION</i> datasets to verify the effectiveness of the proposed decision-making method and in particular, its real-time performance is shown by employing a hardware-in-the-loop (HiL) experiment bench.