Robotic Learning with Reusable Datasets

点击翻译按钮获取中文摘要

One of the major challenges in modern robotics is making robots capable of performing a broad range of tasks in open-world environments, such as homes, offices, hospitals, and construction sites. Such sites are characterized by being very different from each other and include events that are unknown ahead of time. Machine learning has emerged as one of the most effective approaches to allow such broad generalization, particularly in areas that require human like sensing, such as the visual sense. However, effective applications of machine learning to robotics suffer from a major problem: they require collecting large enough observations (referred as to datasets) to enable such broad generalization. While widely reused and shared datasets have enabled broad generalization in areas, such as computer-based vision recognition, this data reuse is difficult in robotics. In this project, the investigators’ goal is to develop methods and techniques that can make it possible to utilize large reusable datasets for robotic learning, such that the same data can be reused repeatedly for a wide range of tasks and domains (with some modest amount of domain-specific collection in each case), while enabling broad generalization. The focus will be on methods for directly learning new skills for robots from the visual observations, using both human-provided data and collected data by the robot itself. The investigators will aim to both develop such algorithms and to collect and disseminate suitable datasets that other researchers can reuse. If successful, this project may lead both to new methods for controlling robots in diverse real-world settings, and tools and resources that can further facilitate future research on robotic learning, making it accessible to scientists and engineers that may not have the capability to collect large datasets on their own.Enabling robotic learning with reusable data requires resolving several important questions: How do people develop robotic learning techniques that can reuse such data? How can people gather the kinds of datasets that can be used for multiple robots, applications, and environments? Answering these questions will require new algorithmic tools for robotic learning, new data collection methodologies, and of course collecting the datasets themselves. The investigators’ technical approach will be focused around two key thrusts. Thrust 1 will be concerned with algorithms development for data reuse, which itself will be divided into two parts: the first part, focuses on imitation learning, and the second part, focuses on reinforcement learning. In Thrust 2, the objective is to collect large and reusable datasets that can be effectively utilized by the broad robotics research community. The first part of Thrust 2 will focus around the development of open-source tools for high-volume data collection. The second part will focus on collecting and disseminating the datasets themselves.This project is supported by the cross-directorate Foundational Research in Robotics program, jointly managed and funded by the Directorates for Engineering (ENG) and Computer and Information Science and Engineering (CISE).This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.