Over the past few years, I have been thinking quite a bit about what I consider to be a central issue at the interface of machine learning and robotics – can we incorporate prior knowledge regarding dynamical behavior into learning algorithms in such a way that it remains invariant with respect to the inevitable quantitative uncertainties that plague real systems? In my own work, I have taken the view that the right approach is to focus on the topological structure of strategies.
I am very pleased to see that this view is becoming more mainstream. There is a workshop at NIPS focusing on ways to learn topology from data. Meanwhile, in the general area of robotics, there is a stellar group of researchers (see, e.g., the SToMP project) studying the use of topological methods to deal with distributed robotics and sensor networks.
Of course, there is still a lot more to be done before these methods begin to show tangible results at the level of something like a humanoid robot. Nonetheless, I am pleased to see this emerging consensus. It’s the right way!