Dr. Arun Singh, Associate Professor, University of Tartu, Estonia
Talk title: End-to-End Learning with Differentiable Optimization
Abstract:Data-driven planning and control with deep neural networks can produce very cool demos. But they also tend to be very brittle, preventing their deployment in safety critical applications. In this talk, I will present our efforts towards making learning based planning and control approaches more deployable by fusing certain algorithmic structure into the neural network architectures. Specifically, I will talk about embedding trajectory optimization as a layer within neural network policies, and present some recent results of ours towards this end, that covers autonomous driving and drone navigation. I will also cover our plans to expand research into this area in a soon-to-start European Union funded project to push AI based planning and control towards deployability.