AI Safety under Distribution Shift | Lars Lindemann (ETH Zürich) | #12

🎓 TU Delft | Delft Center for Systems and Control (DCSC) 🇳🇱 📚 Colloquia Series – Recording #12 How can we guarantee the safety of AI-powered autonomous systems when their own actions change the data they observe? In this DCSC colloquium, we explore safe control under interaction-driven distribution shift, introducing conformal prediction as a powerful tool for uncertainty quantification and showing how robust conformal prediction enables provably safe control in dynamic, interactive environments. 🚀 Welcome to the official recordings of the DCSC Colloquia Series at TU Delft! In this series, we share insightful talks from leading researchers presenting their latest work in systems, control, optimization, and learning. 🔗 https://www.tudelft.nl/me/over/afdeli… ⸻ 🎥 Talk Title: Safety under Interaction-Driven Distribution Shift 📣 Speaker: Lars Lindemann (ETH Zürich) Scholar: https://scholar.google.nl/citations?u... Website: https://sites.google.com/view/larslin... ⸻ 🧠 Abstract: As autonomous systems increasingly rely on AI and data-driven models, ensuring their safe operation in complex and interactive environments has become a fundamental challenge. While statistical learning methods often assume that training and deployment data follow the same distribution, this assumption breaks down when autonomous agents interact with their surroundings and influence the very data they receive. This talk introduces conformal prediction, a modern statistical framework for uncertainty quantification, and demonstrates how it can be used to design controllers with formal safety guarantees. Building on this foundation, Prof. Lindemann presents recent advances in robust conformal prediction, enabling safe control under interaction-driven distribution shifts where the behavior of other agents and the environment continuously changes the underlying data distribution. The framework is motivated by applications in autonomous driving, robotics, wildfire prevention, and disaster response, illustrating how learning-based control can remain both adaptive and provably safe in real-world deployments. ⸻ In this talk: 🛡️ Safe learning-enabled control 📊 Conformal prediction for uncertainty quantification 🔄 Interaction-driven distribution shift 🤖 Autonomous systems and robotics 🚗 Safe autonomous driving 📐 Robust conformal prediction 🌍 AI safety in interactive environments ⸻ Chapters: 00:00 - Introduction and Seminar Overview 01:32 - Motivation: Challenges in Autonomous Systems 05:02 - Problem Formulation: Safe Control in Dynamic Environments 09:53 - Introduction to Conformal Prediction (CP) 15:57 - Understanding Marginal vs. Conditional Coverage 17:15 - CP for Safe Control Design 20:04 - Multi-step Prediction Regions and Normalization 23:54 - Model Predictive Control (MPC) Integration 25:27 - Simulation Results: Mobile Robots and Carla Simulator 26:33 - Safe Control in Interactive Environments 28:53 - Iterative Control Update Scheme for Distribution Shifts 32:01 - Adversarially Robust Conformal Prediction 38:38 - Convergence Conditions and Closed-Loop Analysis 42:10 - Generalizing to Robust Control Design 44:11 - Q&A Session ⸻ 🔬 Research Clusters: Control and Learning • Autonomous Systems ⸻ 👤 About the Speaker: Lars Lindemann is an Assistant Professor at ETH Zürich, where he leads research on the safe and reliable operation of autonomous systems. Before joining ETH Zürich, he was an Assistant Professor at the University of Southern California (USC). He received his PhD from KTH Royal Institute of Technology and completed a postdoctoral appointment at the University of Pennsylvania. His research focuses on safe learning-enabled control, uncertainty quantification, formal methods, and autonomous systems, with applications spanning robotics, autonomous driving, and safety-critical AI. His work has received several distinctions, including the Outstanding Student Paper Award at the IEEE Conference on Decision and Control (CDC). ⸻ 📈 Keywords: conformal prediction, robust conformal prediction, uncertainty quantification, safe control, AI safety, autonomous systems, distribution shift, learning-enabled control, robotics, autonomous driving, control theory, machine learning, safety-critical systems, TU Delft, systems and control, DCSC Hashtags: #ConformalPrediction #AISafety #AutonomousSystems #ControlTheory #MachineLearning #Robotics #AutonomousDriving #ETHZurich #TUDelft #DCSC