Learning Uncertainties for Safe Control: Theory and Applications

Professor Yulong Gao 

 

Uncertainties are inevitable for control of dynamical systems of practical relevance. The safety boundary of the systems is sensitive to the uncertainty size. A fixed and over-approximation of the uncertainty set may lead to conservative solutions. In the first part of this talk, we will present new data-driven methods to learn uncertainty sets in real time, with provable guarantees. We will first show how to embed the learned sets in the tube model predictive control. These learned uncertainty sets can be adaptively used to update the tube, without increasing the online computation complexity. In the second part of the talk, we exploit these methods for safe motion planning in dynamic environments. We develop efficient, robust, and safe motion-planning algorithms that integrate real-time learning of the obstacle uncertainties. These algorithms can naturally capture the control intention in the multi-agent systems. We demonstrate our results in both simulations and hard-ware experiments.