IRFP: Deep Neural Networks for Perception and Action Integration in Robotic Control

Lockett Alan J, Austin TX

Investigators

Abstract

The International Research Fellowship Program enables U.S. scientists and engineers to conduct nine to twenty-four months of research abroad. The program's awards provide opportunities for joint research, and the use of unique or complementary facilities, expertise and experimental conditions abroad. This award will support a twenty-four-month research fellowship by Dr. Alan Lockett to work with Professor Juergen Schmidhuber at the Instituto dalle Molle di Studi sull'Intelligenza Artificiale (IDSIA) in Lugano, Switzerland. This project explores methods for training deep neural networks to control a physical robot with humanoid hands and arms. The recent development of new methods for training deep artificial neural networks has resulted in breakthroughs on a number of benchmarks in artificial intelligence. Deep neural networks currently hold the record for benchmark predictive tasks including handwriting recognition (MNIST) and object recognition (NORB, CIFAR-10). This project is developing methods for training deep neural network controllers with thousands or millions of parameters using an array of massively multiprocessor GPUs with over 4,096 processors. Deep network controllers in this research are trained using neuroevolution, reinforcement learning, and combinations of the two. Such controllers are being used to train a humanoid iCub robot to manipulate objects for the AAAI Small-Scale Manipulation Challenge. The research is being performed at the Instituto dalle Molle di Studi sull'Intelligenza Artificiale (IDSIA) in Lugano, Switzerland in conjunction with Professor Juergen Schmidhuber. IDSIA is a leading research institution in the study of deep neural networks, artificial evolution, reinforcement learning, and robotic control. This research seeks to advance our understanding of robotic control in general. The focus on deep neural networks that integrate hierarchical perception and action modules has the potential to result in breakthroughs in control of complex robotic systems that would enable the deployment of computer and robotic systems that operate with greater autonomy than is currently possible. The deployment of robotic technologies over the course of the next century is likely to mirror the rapid introduction of computer technology in the past century. Behind the success of these robots will be deep hierarchical control systems that integrate perception and action at a high level of abstraction, as studied in this research. This research examines technologies that hold the potential to transform and improve our lives in innumerable ways. In the future, self-driving cars will co-ordinate with each other and with an active roadway to minimize accidents and improve efficiency. Advanced autopilot technology will finally make personal flying vehicles a reality. Autonomous robotic miners will reduce risk to humans while improving access to raw materials and resources. Robotic surgeons will perform complex operations with new levels of precision. Each of these technologies depends critically on the availability of deep integration of perceptual analysis and hierarchical controllers. The use of deep neural networks like the ones studied in the proposed research constitutes a promising approach to bringing these new technologies out of the lab and into our daily lives.