As we read in the newspaper daily about the success and promise of deep learning, I thought I would take this opportunity to wave the flag for the model-based approach. I believe that one of our most fundamental challenges as a field right now is to transition our success stories from YouTube out into the real world. Fielding a humanoid robot in a disaster environment, or flying a UAV at high speeds through a cluttered environment requires reliable online planning in novel environments, and robustness to uncertainty from perception, imperfect actuators, and model errors.
I believe that transitioning dynamic robots to the real world requires an explicit focus on modeling what we know about the world, and explicitly designing in robustness to situations that we have seen yet. These have natural formulations using optimization. Making these optimizations tractable requires exploiting sparsity and convexity in our robot equations, and making informed relaxations. In this talk, I will review our best attempts to date and give examples with fast vision-based UAV flight through clutter and MIT's entry into the DARPA Robotics Challenge.