Event Details

PhD Thesis Proposal - Kegan Strawn

Wed, Apr 30, 2025
11:00 AM - 12:00 PM
Location: GCS 502C
Title: Conformal Prediction for Safe Robot Planning in Dynamic Environments
 
Date and Time: Wednesday, 04/30/25 - 11:00a - 12:00p
 
Location: GCS 502C
 
Committee Members: Lars Lindemann, Nora Ayanian, Jyotirmoy Vinay Deshmukh, Erdem Biyik, Ketan Savla
 
Abstract: Safe robot navigation in dynamic environments around other uncontrolled agents is a central challenge for robotics. This thesis proposal explores statistical tools to quantify uncertainty in control and planning for collision avoidance applications in new and challenging problem settings. First, we introduce conformal predictive safety filters, which augment reinforcement learning policies with learned safety layers that avoid uncertainty regions around dynamic agents, providing probabilistic safety guarantees and reducing collisions without being overly conservative. We then extend this idea to multi-agent pathfinding (MAPF) with CP-Solver, a novel variant of Enhanced Conflict-Based Search that plans around uncontrollable agents. By incorporating uncertainty-aware predictions into planning, CP-Solver offers probabilistic safety guarantees while maintaining high throughput. We conclude with future work on online model selection to robustify and adapt safety filters in real-time, demonstrating safety and performance results through multi-robot drone simulations. Together, these contributions advance safety guarantees and performance in multi-agent systems by combining prediction, uncertainty quantification, and planning.