Risk-Aware Decision-Making and Planning Using Prediction-Guided Strategy Tree for the Uncontrolled Intersections

Uncontrolled intersections with interaction and uncertainties are challenging for autonomous vehicles (AV) to manage. In this work, we propose a decision-making model specific to intersections with emphasis on three aspects. First, behavior estimation of the social vehicles' (SVs) is essential for risk avoidance. We try to improve prediction accuracy by predicting the intentions and driving styles of SVs in advance and doing adaptive goal sampling. Second, the uncertainty from the prediction results should be considered in the decision-making process. For this, a risk-aware framework is developed, composed of a Subordinate Driver (SD) and a Primary Driver (PD) for decision-making and planning. Particularly, in SD, the prediction-guided strategy tree is built to search for an optimal strategy with observation and action branch trimming, which employs the prediction results for risk assessment. In PD, to mimic the both-way negotiation among vehicles, the level-k game model is deployed to determine the action in the players' best interest and update the estimation of driving styles. Third, the generated maneuver is required to be evaluated in a closed-loop simulation. A 'semi-autonomous' control model is designed, which is a combination of the dataset and the stochastic sampling model. The results of ablation experiments verify the function of each module. The case studies and comparison experiments demonstrate the effectiveness of the framework in highly interactive intersections.