Boundary Sampling to Learn Predictive Safety Filters via Pontryagin's Maximum Principle

A team of researchers from UC Berkeley and MIT has published a novel method for training safer autonomous systems. The paper, 'Boundary Sampling to Learn Predictive Safety Filters via Pontryagin's Maximum Principle,' tackles a core problem in AI safety: efficiently gathering the right data. Traditional learning-based safety filters need examples of states likely to cause violations, but randomly sampling complex systems (like autonomous cars) is inefficient and often misses critical edge cases.

The new method applies Pontryagin's Maximum Principle (PMP), a cornerstone of optimal control theory, to mathematically characterize 'boundary trajectories'—paths where the system operates at the very edge of safety without crashing. These trajectories are then used to guide data collection for a learned Hamilton-Jacobi reachability analysis, concentrating computational effort where it matters most. The result is a learned Control Barrier Value Function that acts as a real-time safety filter.

In validation tests, including a shared-control automotive racing simulation, the PMP sampling approach demonstrated major improvements. It achieved faster convergence during training, significantly reduced failure rates, and provided a more accurate reconstruction of the true 'safe set' of states. Crucially, the final safety filter operates with a wall time of around 3 milliseconds, making it suitable for real-time control in high-stakes applications like autonomous driving.