Flexible Vision-Based Robotic Manipulation via Meta Learning and Deep Reinforcement Learning
Abstract
Robots that are autonomous and versatile have the potential to be tremendously useful for a range of manual, tedious, or dangerous tasks that are currently performed entirely by humans. However, in order to be successful in performing manipulation tasks in diverse, dynamic, and unstructured environments, robots must be able to learn continuouslyacross tasks, environments, and even experimental set-upsperpetually accumulating their experiences and learning from them. This setting, however, presents massive challenges for current robot learning approaches, which largely assume carefully-designed experimental set-ups and state representations, narrow singletask learning problems, and supervision in the form of reward feedback or labels. In this project, we aim to explore how we can enable robots to learn continuously from raw perception inputs. In particular, we will address the following questions. Can algorithms leverage large amounts of diverse data, across environments and experiments, to acquire skills that generalize broadly? Can robots leverage large datasets of videos of human behavior to broaden the generalization of behaviors and goal representations? How can algorithms leverage compositional structure to learnincreasingly complex tasks over time? We will study each of these questions in the context of three research themes: first, how visual world models can be learned across diverse, multi-domain data including videos of humans for solving a breadth of tasks; second, how meta-learning can be used to continuous learn a sequence of tasks including a curriculum of compositional tasks; and, third, how robots can develop internal representations of tasks from diverse data sources
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 20, 2020
- Source ID
- N000142012675
Entities
People
- Chelsea Finn
Organizations
- Office of Naval Research
- Stanford University
- United States Navy