Differentiable MPC, Control Barrier Functions and Reinforcement Learning

The intersection of learning-based methods and optimal control has emerged as one of the most active research frontiers in control engineering. Specifically, combining Model Predictive Control (MPC) with Imitation Learning (IL) and Reinforcement Learning (RL) has already shown very promising results in a wide range of applications.
In many of these methods, the central element is the use of MPC as a structured, differentiable policy within IL/RL frameworks. By treating MPC as a differentiable layer, with cost-function parameters dynamically tuned by a neural network and the whole policy trained end-to-end, one can retain the constraint-handling and model-based planning capabilities of MPC while leveraging the flexibility of learned representations, even for high-dimensional data.
At the same time, Control Barrier Functions (CBFs) have become a popular principled mechanism to guarantee safety in MPC. Still, their effectiveness revolves around the tuning of the class-K functions that govern the trade-off between performance and conservativeness, which opens a possibility for dynamically learning these parameters from interactions with the environment.
This talk will go over the basics and current literature on the combination of MPC, IL/RL, and CBFs, including work I did during my PhD on end-to-end autonomous driving through differentiable MPC and IL. Then, it will suggest open questions in this domain, to spark an open discussion with the audience on techniques and approaches we could explore during my research visit at IMT.

Join at imt.lu/aula1