
Towards Safe and Efficient Learning in the Physical World
Andreas Krause of ETH Zurich speaks in Stanford's AA289 Robotics and Autonomous Systems Seminar on how agents can learn safely and efficiently through real-world interaction. He presents safe Bayesian optimization, where uncertainty in an unknown objective and its constraints is quantified to guarantee both safety and convergence to a reachable optimum, then moves to Bayesian model-based deep reinforcement learning, where epistemic uncertainty in a learned world model guides exploration while keeping the system safe. He closes with meta-learning of flexible probabilistic models from related tasks and simulations, illustrated with robotics applications and the tuning of the SwissFEL Free Electron Laser. The talk is technical, aimed at a research audience familiar with Bayesian methods and reinforcement learning, and grounds its theoretical guarantees in concrete engineering problems rather than toy examples.