Interactive prediction and decision-making for autonomous vehicles: online active learning with traffic entropy minimization

Abstract

Interacting with the surrounding road users is crucial for autonomous vehicles (AV). However, the inherent multimodality and uncertainties associated with traffic participants (TP) pose challenges in AVs' prediction and decision-making (PnD). A primary challenge is adapting predictors trained on static offline datasets to the dynamic, diverse data streams encountered in reality. Secondly, utilizing one single forecast trajectory with the highest probability for decision-making contains potential risks as it neglects that even a small probability represents a subset of TP behaviors. Based on the existing prediction backbone, we propose an online learning approach incorporating pseudo-labels inferred from partial feedback as compensation for conventional methodologies, considering both the commonsense and personalization facets of driving. Drawing inspiration from the second law of thermodynamics, we propose to minimize microscopic traffic entropy as an additional objective in decision-making. This objective aims to reduce the chaos of traffic scenes, thus achieving more predictable future interactions and, conversely, making future decisions easier. Through real-time human-in-the-loop experiments, we quantifiably and comparably reveal that adopting one single trajectory without online learning in PnD is risky. However, this reliability is verified to be significantly improved by our proposed techniques, and the efficacy is further analyzed in a subsequent qualitative study. A static experiment transferring the prediction algorithm trained exclusively on Argoverse 2 to datasets including NGSIM, HighD, RounD, and NuScenes is also conducted, demonstrating that the proposed correction can effectively mitigate the gap between the datasets and real-world scenarios.