Learning for robotics
Many problems in robotics have unknown, stochastic, high-dimensional and highly non-linear dynamics, and offer significant challenges to existing algorithms. Some of the key difficulties in these problems are that (i) It is often hard to write down, in closed form, a formal specification of the control task (for example, what is the objective function for "flying well"?), (ii) It is difficult to build a good dynamics model because of both data collection and data modeling challenges (similar to the "exploration problem" in reinforcement learning), and (iii) It is expensive to find closed-loop controllers for high dimensional, stochastic domains.
Our research focuses on developing novel algorithms that leverage learning techniques to tackle challenging real-world control tasks. We are interested in learning in a variety of settings, including learning from demonstrations, learning through more general forms of expert advice/supervison, and learning through safe autonomous explorations.
We have successfully applied such techniques to enable a quadruped robot to traverse challenging, previously unseen terrains, and a helicopter to perform by far the most challenging aerobatic maneuvers performed by any autonomous helicopter to date, including maneuvers such as chaos and tic-tocs, which only exceptional expert human pilots can fly.