[Hybrid Seminar] "Towards Hierarchical Motor Control" by Dr. Steve Heim, Biomimetic Robotics Lab at MIT

Meeting URL:   https://oist.zoom.us/j/97352227518?pwd=QTVqRkFaSTE3VHdFdWNKdGlnRGVMQT09&from=addon

Meeting ID: 973 5222 7518

Passcode:   629035

Speaker: Dr. Steve Heim, Postdoc, Biomimetic Robotics Lab at MIT

Title: Towards Hierarchical Motor Control

Abstract:   How do we learn new skills like parkour or dance moves? Agile motions require substantial practice; however, we are also adept at guiding our own learning process. We visualize the motions beforehand, practice similar but simpler motions to build confidence, creatively combine previously learned skills to invent new moves, and can switch learning strategy if we get stuck.
We believe learning agents need to be able to reason by making predictions at different time-horizons and resolutions, which can be adjusted according to the learning task, and their current confidence level. In other words, we need to develop a flexible hierarchy of predictive models, policies, and reward functions.
I will briefly introduce the basic framework we're currently working towards and some preliminary results. I will also introduce viability measures, as a means to estimate "confidence" of a policy.