SLES: Certified Learning of Safety Certificates for Learning-Enabled Dynamic Systems

Robotic navigation and control is reaching new levels with the introduction of neural network-based control mechanisms. Trained largely in simulation environments, these learning-based controllers have been used to guide a kitchen robot, enable a robotic goalie to deflect soccer balls, and control a two-legged robot to jump to new heights and self-stabilize on landing.

It is now a reality that robots can help with the decades-old problems of learning to do a complex task, and learning how to interact with a complex environment. However, the more complex the task and environment, the harder it is to ensure that the system will operate safely. Automation in the real world requires a characterization of the safety of such systems.

One popular way to do this is through the use of safety certificates (e.g., using mathematical approaches that are guaranteed). The project’s novelties are in developing methods that use deep learning to develop safety certificates in realistic dimensions. The project’s impacts are in the safe design of systems which use learning-enabled perception and prediction modules in an autonomous system.

For such systems, it is crucial to understand and predict the behavior of the autonomous system with the learning-enabled components. The project is training the next generation of computer scientists and engineers in these tools and methods, and is further developing a new course at Berkeley in integrated perception, learning, and control.

This project addresses two key objectives: (I) Learning safety certificates and their control policies and (II) certifying the learned system. The project is developing new techniques to certify the system with learning-enabled components, employing techniques which propagate uncertainty through system modules. While computing safety certificates has required solving a Hamilton-Jacobi equation, or devising a control barrier function, these methods have been restricted to low dimensional problems.

The project is using learning-based methods to compute safety certificates and control policies at scale. To certify the learned system, the project is developing methods for out-of-distribution data detection and management, techniques to characterize uncertainty from learning-enabled components and using it to assess the safety of the learning-enabled system, and tools which take data-specific metrics into account, like Lyapunov density models.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.