Safety-Dominant Stochastic Model Predictive Decision-Making Considering Obstacle Trajectory Uncertainties for Intelligent Vehicles

Robotaxis tend to hesitate when encountering obstacles with uncertain future trajectories because their decision-making strategies are overly cautious, aiming to ensure robust safety. To reduce conservatism while ensuring reliable obstacle avoidance, this paper introduces Stochastic Model Predictive Decision-Making (SMPDM), considering the probabilistic distributions of obstacle future trajectories. Firstly, an integrated longitudinal and lateral stochastic decision-making model is established, utilizing real-vehicle data to model the uncertainty distribution of modeling errors. Secondly, a collision risk assessment method based on the Separating Axis Theorem (SAT) and Monte Carlo (MC) simulations is proposed. This method accounts for both the geometric dimensions of the intelligent vehicle and obstacles, as well as the time-varying uncertainty in their future trajectories. Thirdly, a safety-dominant dual decision-making optimization problem was formulated for minimizing collision risks under high-risk conditions. This dual optimization problem was further transformed into a piecewise optimization problem, with a proof of equivalence provided for the transformation process. The real-time solution of this stochastic model predictive optimization problem was realized through GPU-based parallel computing. Finally, utilizing the Next Generation Simulation (NGSIM) dataset, we conducted Hardware-in-the-Loop (HiL) experiments on an industrial computer, performing both ablation studies and comparative analyses. Additionally, we conducted real vehicle experiments. The result indicates that, by establishing stochastic models that consider uncertainties in future trajectories and designing safety-dominant decision-making approaches, our method ensures safety while reducing vehicle conservatism.