Stanford Course Overview: Validation of Safety Critical Systems (AA228V)

Autonomous systems can't be trusted in the real world until you've found how they fail. In this course overview video for Validation of Safety Critical Systems (AA228V), Stanford's Sydney Katz, Lecturer, and Mykel Kochenderfer, Associate Professor, explain how. Validation of Safety Critical Systems (AA228V) is open now for winter quarter enrollment: https://stanford.io/AA228V In this course, you'll learn to: • Use simulation to find failures in autonomous systems before deployment • Estimate a system's probability of failure with sampling-based methods such as importance sampling and Markov chain Monte Carlo • Specify safety properties with temporal logic and verify them with formal methods such as reachability analysis Commercial aviation is one of the safest places you can be, and Kochenderfer credits decades of disciplined verification and validation. In this conversation, he and Katz cover why that playbook isn't enough for the next generation of autonomy: today's AI systems have too many variables to test the old way. Both Stanford faculty members make the central claim that you have to understand how a system behaves before it goes out into the world. They show how hard it is to specify what you actually want with a wildfire example. A simulated fleet of firefighting drones was rewarded for every grid cell of fire extinguished, so the drones stayed away from the fire, let it burn across the entire map, and then put out as many cells as possible. Keep watching to understand why the objective looked perfectly sensible, but the optimal solution was exactly what no one wanted. Browse all available graduate courses: https://stanford.io/Winter-2026 #AISafety #AutonomousSystems #SafetyCriticalSystems