Risk-Controllable Autonomy: A Principled Framework from Perception to Action with Conformal Guarantees

Autonomous systems, from self-driving cars to industrial robots, are poised to revolutionize society, yet their widespread deployment is critically hindered by the challenge of uncertainty. While deep learning models excel in controlled environments,their potential for rare but catastrophic failures in the unpredictable real world makes them untrustworthy for safety-critical applications. This creates an urgent need for advanced artificial intelligence systems that do not just perform well on average, but can provide rigorous, provable guarantees on their safety. Current research often addresses uncertainty in isolated components of the autonomy stack: perception, prediction, or decision-making, leaving a critical blank in creating an integrated, certifiably safe pipeline.This project introduces a unified, principled framework that leverages advanced Conformal Prediction (CP) theory to provide rigorous, distribution-free safety guarantees across the entire autonomy stack. We directly confront three fundamental challenges. First, to move beyond perception models for semantic segmentation with only "average-case" guarantees, we will develop novel methods that achieve more robust conditional risk control, adapting their uncertainty estimates to the difficulty of each specific scene. Second, to address the multi-modal nature of agent behavior, we will design a novel trajectory prediction method that generates planner-friendly geometric sets such as convex polytopes and alpha-shapes that are simultaneously guaranteed to be valid in coverage rate, geometrically simple for downstream planning use, and tight enough to be efficient and practical. Third, and most critically, we will move from guarantees to action by designing an online risk controller. This controller operationalizes Conformal Decision Theory (CDT) in a closed-loop system, dynamically adjusting the planner's conservatism based on realized outcomes to provably steer the system's long-term average safety performance to a user-defined target, ensuring robustness even under distribution shifts.By creating a principled and complete pipeline, this project marks a fundamental shift from static uncertainty avoidance to dynamic, adaptive risk control. The outcomes will provide a foundational methodology for building the next generation of trustworthy autonomous systems, accelerating their safe and reliable deployment in society.