PhD Proposal: Neurocomputational Cause-Effect Reasoning for Imitation Learning in Cognitive Robots

Talk
Garrett Katz
Time: 
11.23.2015 14:00 to 15:30
Location: 

AVW 3258

Robots with advanced cognitive abilities have the potential to deliver great benefit to humankind. To make such robots useful on a large scale, it is important that they can be taught new skills in an intuitive way by humans without technical expertise in robotics. A promising route in this direction is imitation learning, whereby a robot learns new skills by observing human demonstrations. However, cognitive-level "imitation learning" today is generally confined to highly constrained task scenarios that rely on hand-crafted, idealized representations of the world. This limits the robot's ability to generalize learned skills to new situations.
This proposal holds that cause-effect reasoning is a crucial element of generalization in cognitive- level imitation learning, and that implementing cause-effect reasoning with brain-inspired computation will reduce a robot's reliance on hand-crafted representations and improve its ability to generalize. A framework for imitation learning is presented in which cause-effect reasoning is used to infer the high-level goals of a demonstrator, enabling better generalization to new situations where the low-level details are different. A traditional, non-neural implementation is validated on board a dexterous robot that learns skills related to assembly and maintenance. Preliminary work towards a neurocomputational implementation of this framework is also presented. Initial results suggest that this framework constitutes a promising approach to the problem of cognitive-level imitation learning. A research plan is devised to develop this proof of concept into a general-purpose, highly effective, and rigorously evaluated imitation learning system.
Examining Committee:
Committee Chair: - Dr. James A. Reggia
Dept's Rep. - Dr. Howard Elman
Committee Member: - Dr. Dana Nau