Research Preparation Exam: Pei-Chi Huang

Date: July 10, 2013
Time: 1pm - 2pm
Place: GDC 6.516
Committee members: Peter Stone, Risto Miikkulainen, and Aloysius K. Mok(Chair)

Title: Cyberphysical Robotic Hand Grasping through Neuroevolution

Abstract:
An ambitious goal of cyberphysical avatar research is to enable robots operating in an unknown environment to successfully perform arbitrary physical tasks in real-time given only high-level human supervision. However, many such low-level physical tasks like object grasping cannot yet be fully automated. In particular, grasping a target object without advance knowledge of its three-dimensional model remains challenging; yet such models are not always readily available. Thus, motivated by the promise of neural networks for feature extraction, this paper explores neuroevolution, i.e. evolving neural networks with evolutionary algorithms, as a means for object grasping without relying upon pre-computed 3D object models. Given only depth data from a Microsoft Kinect sensor, the approach in this paper evolves neural networks to identify the appropriate hand positions and orientations for grasping objects. The advantage of such an approach is that once trained off-line, a neural network requires little computation to execute and no additional on-line learning, enabling real-time applications. After training a set of grasping neural networks in many simulated scenarios, the most promising one can then be applied to control object grasping in a real cyberphysical avatar.