Скачать или смотреть Stanford Seminar - Continual Safety Assurances for Learning-Enabled Robotic Systems

November 15, 2024
Somil Bansal, Stanford University

The ability of machine learning techniques to leverage data and process rich perceptual inputs (e.g., vision) makes them highly appealing for use in robotic systems. However, the inclusion of data and machine learning in the control loop poses an important challenge: how can we guarantee the safety of such systems? To address these safety challenges, we present a controller synthesis technique based on the computation of reachable sets, using optimal control and game theory. We present new methods that leverage advances in physics-informed neural networks to compute reachable sets and safety controllers efficiently. These techniques are highly scalable to high-dimensional systems, enabling the learning of safe controllers for a wide array of robotic systems. Furthermore, these methods allow us to dynamically update safety assurances online, as new environment information is obtained during deployment. In the second part of the talk, we will present a toolbox of methods that use data-driven reachable sets to stress-test learning and vision-based controllers. By identifying safety-critical failures, these tools guide performance improvement while maintaining safety. Together, these advances establish a continual safety assurance framework for learning-enabled robotic systems, where safety considerations are integrated across various stages of the learning process, from initial design to deployment and ongoing system enhancement. Throughout the talk, we will illustrate these methods on various safety-critical autonomous systems, including autonomous aircrafts, autonomous driving, quadrupeds, and quadcopters.

About the speaker: https://smlbansal.github.io/

More about the course can be found here: https://stanfordasl.github.io/robotic...

View the entire AA289 Stanford Robotics and Autonomous Systems Seminar playlist:    • Stanford AA289/ENGR319 - Robotics and Auto...  

► Check out the entire catalog of courses and programs available through Stanford Online: https://online.stanford.edu/explore