The next IPAB Workshop will take place on 17/09/2020 at 1pm via Blackboard Collaborate. Please refer to your emails for the link.
Speaker: Daniel Angelov
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.