Sensorimotor Representation Learning for Robotic Motion Skills
Abstract
Technical Approach. Some of the most effective algorithms for learning representations come from the field of deep learning. However, extending deep learning to sensorimotor tasks in robotic control has proven to be challenging. Deep learning typically requires large amounts of training data (e.g., 1.3 million images for object recognition). Robotic sensorimotor learning requires data from task-specific robotic interactions, which are expensive and time-consuming, and it is unclear whether the same kinds of hierarchies that have been effective in vision will produce good results for control. The PI proposes to explore these challenges through three research tasks:Task 1: Semisupervised Action-Perception Representations Learning. The PI proposes to investigate semisupervised learning methods that will allow sensorimotor representation learning toincorporate large prior datasets of images or motor commands. While these prior datasets do not by themselves describe the sensorimotor relationship, they can inform the algorithm about the structure of each modality, allowing it to learn substantially more generalizable representations.Task 2: Multi-Robot Learning and Lifelong Learning. The PI also proposes to augment the datasets available for learning sensorimotor representations by simultaneous learning of robotic skills on multiple robots, and combining experience from multiple tasks over a robot~s lifetime.Task 3: Action-Centric Hierarchical Representation Learning. The PI will explore how hierarchicalrepresentation learning algorithms can be designed so as to explicitly couple perception and action, and to learn hierarchies of arbitrary depth, acquiring progressively more abstract concepts. These research questions will be discussed in the context of challenging robotic applications, including rapid autonomous learning of assembly tasks, handling complex objects such as dishes in a personal robotics application, and obstacle avoidance for autonomous aerial vehicles.Anticipated Outcome. The proposed research will advance the state of the art in robotic learning,enabling autonomous acquisition of sensorimotor skills that generalize more widely due to powerful, generalizable, learned representations. Impact on DoN Capabilities. Autonomous acquisition of generalizable sensorimotor skills provides for a range of applications with high DoN impact. Examples include autonomous repair and maintenance of equipment in the field, automation of warehousing, and a variety of otherautonomous robotic missions in unstructured natural environments.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 04, 2017
- Source ID
- N000141712118
Entities
People
- Sergey Levine
Organizations
- Office of Naval Research
- United States Navy
- University of California Regents