Events

Safety Assurances for Learning-Enabled Robotic Systems

Speaker: Somil Bansal

Location: 370 Jay Street, Room 825

Date: Wednesday, March 27, 2024

The ability of machine learning techniques to leverage data and process rich sensory inputs (e.g., vision) makes them highly appealing for use in robotic systems. However, the inclusion of learning-based components in the control loop poses an important challenge: how can we guarantee the safety of such systems?

Control theory provides a number of powerful methods for designing safe controllers for robotic systems, such as Hamilton-Jacobi (HJ) reachability analysis. However, these methods often lack the scalability and flexibility to interface with real-world data and machine-learning models. We present neural reachable tubes that leverage traditional HJ safety conditions within a machine-learning framework to design safe controllers from data. Neural reachable tubes are easily scalable to high-dimensional systems, allowing us to learn safe controllers for a broad range of robotic systems. We will next present a toolbox of methods that can leverage neural tubes to update the safety guarantees online within a fraction of milliseconds as new environment information is obtained during deployment. In the second part of the talk, we will discuss how we can use neural tubes to stress-test learning and vision-based controllers to discover their safety-critical failures and use them to improve the controller.

Together, these advances provide a continual safety framework for learning-enabled robotic systems, where safety is integrated in different stages of the learning process, starting from their design to their deployment to iteratively improving the system's safety over its lifecycle. Throughout the talk, we will illustrate these methods on various robotic platforms, including autonomous driving, legged locomotion, and navigating in a priori unknown environments.