ABSTRACT:

We present three approaches to combine formal methods, control theory, and machine learning for safe and performant autonomous systems. 

1. Safe control for learning-enabled systems: We present our recent progress on how to learn safe adaptive behavior for highly interactive multi-agent systems. We will introduce how to quantify the uncertainty of closed-loop control systems using a frequentist method called conformal prediction and incorporate the uncertainty for safe perception-based control.
2. Learning Introspective Control: Oftentimes the systems that we control operate under different conditions due to changing environments, varying system parameters or changes in payload. As such, we strive to develop computationally efficient, data-driven system models that allow predictive controllers to adapt to changes in the environment in real-time. We focus on using Gaussian Processes as models to study the problem in the context of driving on surfaces with changing friction coefficients.
3. Differentiable Predictive Control: Finally, we discuss the application of differentiable predictive control for large-scale urban road networks.