Learning from Humans: Beyond Classical Imitation Learning

May 4, 2022, 1:00 pm2:30 pm
103 Bendheim House (26 Prospect Street) , see abstract for Zoom Link
Event Description

Humans have deep expertise in an incredible array of skill sets. Increasingly robots are being trained to assist in performing some of these tasks, however, robot learning is often limited in two major ways. First, increasingly complex tasks limit the human’s ability to explicitly define task success. Second, modern methods of learning by imitation require a prohibitively large number of samples. We show two novel methods of imitation learning that proveably reduce the number of demonstrations needed to learn a given task. When all we have is a set of offline demonstrations, we show how to bootstrap that data via simulation in order to recover in novel states. If the expert is available during the training process, we show how a new method of learning from interventions unlocks a flexible framework for learning from all types of human interaction.

Zoom link is: https://princeton.zoom.us/j/92823676402