Learning-based planning and control for robotics promises to handle complex dynamics, sensors, and tasks with reduced engineering effort. However, many impressive empirical results are brittle: both the learned policy and the learning algorithm itself may fail under seemingly minor changes in the setting. To realize the full potential of robot learning, we must make these algorithms reliable. In this talk, I will share work towards this goal in both theory and practice.

First, I will present M-GAPS, a principled and general method for optimizing a control policy when both dynamics and costs are time-varying and revealed online. In a broad class of systems with "contractive" dynamics, M-GAPS enjoys guarantees of local or global regret-optimality. In more complex settings beyond the current scope of our theory, M-GAPS remains well-defined; we show strong empirical performance in hardware experiments with quadrotors and cars.

Next, I will demonstrate the benefits of combining the expressive power of deep dynamics models with the interpretability of classic nonlinear control for 1) complex partially-observable tasks such as deformable object manipulation, and 2) off-road driving on challenging terrains using visual foundation models.

Finally, I will outline research plans towards a unified framework for reliable robot learning. These plans center on investigating the structural properties that make learning-to-control tractable in physical systems, while keeping a tight connection between theory and practice.