From Prediction to Planning: Comprehensive Uncertainty Management in Autonomous Driving

Autonomous driving systems face significant challenges in navigating complex, dynamic environments rife with uncertainty. This study proposes a comprehensive uncertainty management framework for prediction-planning systems, which simultaneously models three key uncertainties: short-term aleatoric uncertainty (SAU), long-term aleatoric uncertainty (LAU), and epistemic uncertainty (EU). Leveraging Gaussian mixture models (GMMs) to capture SAU and LAU and deep ensemble techniques to estimate EU, the framework enables concurrent quantification of all uncertainty types. Then, a Comprehensive Uncertainty-Aware Planning (CUAP) approach is developed, integrating customized risk models and a two-stage training strategy to enhance predictive reliability. By establishing a unified foundation for decision-making, the framework addresses the limitations of traditional methods that handle uncertainties in isolation. The contributions include a thorough investigation of uncertainty estimation techniques, risk modeling, and planning strategies, validated through rigorous evaluations using the CommonRoad benchmark and perception-limited scenarios. Results demonstrate substantial improvements over current approaches, particularly in diverse traffic conditions, with the framework enhancing planning accuracy and reliability by integrating multiple uncertainties. This study offers detailed insights into applying uncertainty management from prediction to planning, highlighting its potential to significantly improve autonomous driving performance, especially in accident prevention, through comprehensive uncertainty integration.