Seminar by Dr. Chandana Paul
Learning to Walk: Neural Net Lesions to Test Learning Algorithms
Dr. Chandana Paul
Department of Informatics
University of Zurich
Zurich, Switzerland
Date: Monday, February 23, 2004
Time: 5:00 PM
Venue: CS-103
Abstract
Impelled by a need to plan post-traumatic physiotherapy for patients who have lost their ability for coordinated walking, this research first builds simulated systems that learn how to walk using a GA-optimized Neural Nets. Subsequently, lesions are simulated in this Neural Net to replicate the loss of various sensors and control processes. Based on these results, the contact sensing at the foot is found to be most critical. Subsequently, a minimal walking device was designed with only two control outputs, and is found to be capable of stable walk with a very simple control routine. A number of actual walking robots were then constructed to demonstrate the minimal-DOF designs. Results will be presented from a series of robots real and simulated, with and without upper torsos, which demonstrate differing degrees of competent walking behaviour.
About the Speaker
Dr Chandana Paul holds dual Bachelor's Degrees in CS and in Brain and Cognition, from MIT. She has just finished her PhD at U. Zurich and will be joining Cornell U later in April.