Learning robot behaviors from large data sets is an important way to make robots more capable and reliable. This project will develop algorithms for autonomous robotic skill learning that can easily be used by novice hobbyists with low-cost robots. If deployed widely, such an approach could be used to gather a large number of robotic motions, which can be combined to improve the robot's skills. Availability of large datasets has proven critical in machine learning application areas, from computer vision to speech recognition, and the ability to collect a large amount of robotic interaction data would substantially increase the capabilities of learning-based robotic systems. Since the approach will be designed for untrained users, it also doubles as an effective tool for robotics education.
Deep learning has emerged as a powerful technique for taming the complexity of the real world. The success of deep learning depends on the availability of large datasets, which traditionally have been difficult to obtain for robotic learning. This project will focus on deep learning algorithms that can be used for effective and reliable robotic skill learning, generating intelligent actions directly from raw sensory input, with an eye towards enabling widespread deployment for large-scale data collection. To that end, the proposed research will aim to: (1) devise reliable and robust real-world robotic learning algorithms that can collect experience without human oversight or intervention; (2) build algorithms centered around transfer learning, whereby experience from prior tasks can be used to inform dramatically faster learning of new skills with potentially different robotic platforms; and (3) devise algorithms that can effectively control heterogeneous, low-cost, imprecise robots, so as to facilitate widespread deployment and the project's educational mission.
