Scientists design safety controller with learning capabilities for systems in unknown environments

To ensure that various safety-critical systems can operate safely and work as needed, researchers at Michigan State University have developed a method to design safety controllers with learning capabilities for systems operating in unknown environments, foreign media reported.

Image source: IEEE/CAA Journal of Automatica Sinica

The study, conducted by Bahare Kiumarsi and Zahra Marvi of Michigan State University, focused on the safe operation of autonomous vehicles in urban areas.

Over time, scientists have successfully designed safety control methods based on control barrier functions (CBFs), which can be applied in many fields, including adaptive cruise control, robot safety control, and collision-free multi-agent systems. These methods usually combine CBFs and Lyapunov functions to prove the safety and stability of the controller. Researchers at Michigan State University proposed a novel learning-enabled zero control barrier function (ZCBF) that can operate safely while learning, even if there are unknown dynamics in the environment.

As more and more safety-critical systems are deployed in the real world, scientists must be able to ensure that these systems remain safe at all times. Uncertainties in the environment can affect the safe operation of the system. When studying autonomous vehicles in urban environments, researchers must account for these uncertainties. For example, autonomous vehicles, semi-autonomous vehicles, human-driven vehicles, and pedestrians may appear in the same area.

Therefore, scientists must design a controller that ensures the system operates safely despite uncertainty about how other vehicles and humans navigating the same space will behave. Scientists need to be able to rely on the safety of the system while also making it operate as well as possible. The Michigan State University researchers' approach addresses the problem of how to design safety controllers with learning capabilities for systems that must operate in uncertain environments. Their novel learning method is able to keep self-driving cars safe even when operating with uncertain behavior of other vehicles on the road.

Existing safety control methods require scientists to have a complete understanding of the safe set. Designing safety controls for systems becomes more challenging when there is uncertainty in the environment. These safety-critical systems must be able to quickly learn about uncertainty while achieving maximum safety performance. Slow model learning methods can provide the required safety capabilities but fail to achieve the expected performance. Simple model learning methods based on minimizing modeling error also fail to achieve the required safety requirements, even though the expected estimation error decreases over time. "New learning algorithms are needed to avoid misrepresentation of the safe set as much as possible," said doctoral candidate Zahra Marvi. The method developed by the researchers is able to quickly learn about uncertainty in the environment and quickly achieve safety performance.

"Satisfying safety constraints is critical and needs to be considered during the control design phase, otherwise serious consequences may occur," said Assistant Professor Bahare Kiumarsi. "By using this approach designed by the MSU researchers, the controller in the system can take less conservative actions, resulting in better performance, which in turn improves the safety and performance of the system."