On Uncertainty-Aware Perception, Prediction, and Planning

Urban environments are inherently dynamic and multifaceted, creating conditions of complexity that challenge traditional approaches to sensing, forecasting, and decision-making. In response, artificial intelligence (AI) is increasingly used to power urban systems, enabling them to interpret sensor data, anticipate future states, and plan appropriate actions. A common structure for such AI-driven processes is the Perception–Prediction–Planning (PPP) pipeline, which decomposes complex urban tasks into sequential stages. However, despite its growing role, AI, particularly in the form of deep learning, remains fundamentally limited in its ability to handle uncertainty. Deep learning models often rely on incomplete data, exhibit miscalibrated confidence, and lack mechanisms to enforce logical consistency, allowing uncertainty to accumulate and propagate through the pipeline. Inadequate uncertainty handling, spanning modeling, quantification, and reduction, can result in overconfident yet flawed predictions. Existing uncertainty estimation methods are primarily statistical and frequently overlook the structured relationships and domain constraints that characterize urban environments. Incorporating explicit reasoning mechanisms, such as logical rules or semantic dependencies, refines uncertainty estimates, constrains outputs to plausible interpretations, and ensures coherence across tasks. This dissertation proposes approaches that integrate domain knowledge and symbolic logic to handle uncertainty at each stage of the PPP pipeline, thereby yielding outputs that are calibrated in their confidence and verifiable in their semantics. The contributions are outlined in Chapter 1.4, with detailed summaries of research approaches and key findings presented in Chapters 4 to 7, while the full publications are provided in Chapters P1 to P7. First, we introduce Knowledge-Refined Prediction Sets (KRPS), which embed domain-aware logical constraints into conformal prediction for multi-task perception, ensuring semantically consistent and probabilistically rigorous outputs. Experimental results on the ROAD and Waymo/ROAD++ datasets demonstrate that KRPS outperforms state-of-the-art conformal prediction methods by reducing uncertainty by up to 80% and boosting semantic consistency by up to 30%, all while retaining formal coverage guarantees. Second, we propose NeSyMoF (Neuro-Symbolic Model for Motion Forecasting), a deep learning model, which fuses deep learning with symbolic logic to generate multi-modal trajectory forecasts along with their symbolic rules. Evaluations on Argoverse reveal that NeSyMoF not only outperforms state-of-the-art interpretable baselines for single-mode predictions but also provides first-order logic explanations for its forecasts, articulating the reasoning behind each predicted path. We develop an approach to quantify the uncertainty of NeSyMoF and devise a novel active learning strategy leveraging the discrepancies between the learned rules and the generated paths, which prioritizes high-uncertainty scenarios during training. With our active learning strategy, we reach 99% of the model performance with only 60% of the data. Third, we present SafePath, a path planning approach that combines Large Language Models, conformal prediction for uncertainty handling, and domain-specific safety constraints. We prove theoretically that SafePath limits collision risk to a user-defined probability. Experiments on the nuScenes dataset and the Highway-env simulator show a 77% reduction in planning uncertainty, a 95% decrease in user interventions, and up to a 70% reduction in collision rates compared to state-of-the-art approaches. Finally, we conceptualize and evaluate novel approaches for AI systems to handle uncertainty through interactions with their external environment, with a particular focus on human stakeholders. We investigate this across three key aspects: knowledge transfer and data annotation, where we introduce PointCloudLab, an immersive platform that enhances labeling efficiency and accuracy; communication of reasoning uncertainty, where we develop AR-CP, an augmented reality-based system for conveying uncertainty in perception models; and communication of decision uncertainty, where we explore AR-based representations of planned trajectory uncertainty to improve human trust and decision-making in autonomous vehicles.