IPAB Workshop-17/09/2020

Title: Composing Diverse Policies for Robot Control Abstract: Solving long-horizon problems is a challenging task, requiring  optimizing across a variety of sub-task dynamics. Learning from  demonstration provides a strategy for learning a method to compose a  set of already known diverse controllers, each tuned for their  corresponding sub-problem. Additionally, performing causal analysis on  the controllers gives us the ability to extract specifications about  the demonstrated policy in regards to known symbols in the  environment. Finally, to build robust controllers from demonstrations,  we want to obtain a variety of possible trajectories, often limited by  the comfort space of the demonstrator or the robot. We will present  Learning from Inverse Intervention. A strategy for collaborative  demonstration, in which the robot augments the demonstrator  trajectory, pushing it to uncertain to the policy states. It results  in better demonstrations, as well as the ability to elicit problem  structure.