Stanford AA289/ENGR319 - Robotics and Autonomous Systems Seminar - Stanford Seminar - Safe and Robust Perception-Based Control

Sarah Dean UC Berkeley February 21, 2020 Machine learning provides a promising path to distill information from high dimensional sensors like cameras -- a fact that often serves as motivation for merging learning with control. This talk aims to provide rigorous guarantees for systems with such learned perception components in closed-loop. Our approach is comprised of characterizing uncertainty in perception and then designing a robust controller to account for these errors. We use a framework which handles uncertainties in an explicit way, allowing us to provide performance guarantees and illustrate how trade-offs arise from limitations of the training data. Throughout, I will motivate this work with the example of autonomous vehicles, including both simulated experiments and an implementation on a 1/10 scale autonomous car. Joint work with Aurelia Guy, Nikolai Matni, Ben Recht, Rohan Sinha, and Vickie Ye. View the full playlist: https://www.youtube.com/playlist?list=PLoROMvodv4rMeercb-kvGLUrOq4HR6BZD 0:00 Introduction 0:12 Safe and Robust Perception Based Control 0:46 Machine learning is a promising tool for 1:20 acting on complex information 3:20 Example: Racing from pixels 5:38 Tasks modelled as optimal control problems 6:36 Perception-based optimal control problem 8:26 Perception as virtual sensor 9:28 Problem setting: linear optimal control • Linear dynamics • Complex observation model • Perception map os virtual 10:22 Linear output feedback control Familiar setting 11:05 Aside on low-level control 13:27 Perception: errors and safe set 14:15 Learning and generalization 17:32 Deterministic (adversarial) generalization • Closed loop states depend on the errors 19:02 Closeness implies generalization 20:35 Robust control 22:07 System level synthesis for state feedback 23:56 Optimal control reformulation 29:11 Paradigm for analyzing learning in control 31:06 Output feedback SLS 34:37 Example: static filter and state feedback In the case of a controller of the form 38:47 Robust reference tracking Model waypoints os disturbances 40:23 Example: necessary and sufficient For a simple double integrator example with LOG control, the constraint is necessary and sufficient for stability at origin 41:12 Simulation setting Simplified driving example using CARLA simulator with 2D double integrator dynamics 42:40 Simulation Results 44:45 Real-world demo 44:55 Iterative racing on arbitrary tracks 49:35 Conclusion