Intelligent Robots that Learn to Perform Novel Tasks from Instructions and Demonstrations in Open Worlds
Abstract
Intelligent Robots that Learn to Perform Novel Tasks from Instructions and Demonstrations in Open Worlds.Our overarching goal is a novel synergy of symbolic and subsymbolic approaches and an embodiment in a state-of-the-art humanoid robot platform that will enable genuinely useful cooperative, mixed-initiatve human-robot teams. Combining the power of high-level symbolic representations and reasoning with the flexibility of statistical machine learning of non-symbolicrepresentations situates this project squarely in BAA Topic 1: Integration of Domain Knowledge and Machine Learning. We plan to develop an integrated humanoid robotic system based on the DRACO robot and the DIARC robot control architecture that will be (1) capable of navigating human environments at human speeds and performing various tasks in those environments thatrequire full-body control and manipulation, (2) capable of quickly learning online from observations, demonstrations, and instructions, and applying the learned knowledge immediately during task performance, and (3) capable of learning and using mental models of human teammates to improve task-based interactions with humans. The first goal will allow robots to be deployedin human environments without the need to instrument or otherwise alter those environments for robotic operation; the second goal will allow robots to pick up any knowledge for task performance they do not have on the fly and share it with other robots; and the third goal will allow robots to perform genuinely useful tasks as part of mixed-initiative human-robot teams effectively in a waythat humans find intuitive.The project addresses Focus Area 2 (~Reconcile divergent representations encoded in KR and those that emerge from DL...~) as well as Focus Area 3 (~Develop methods for automatic learning of KR-like structured models of knowledge...~), among others. It has enormous potential for the Navy for tasks such as autonomous shipboard maintenance, emergency response and firefighting, buildig clearing, EOD, or package handling for logistics operations which require nimble and reliable robotic locomotion and manipulation in open environments. The ability to handle the unstructured real world is essential because it will make the robot~s behavior much more adaptive and reliable. Moreover, learning new tasks quickly from single instructions, demonstrations, and observations in open worlds will allow for fast adaptation to new contexts and thus enable the deployment of robotic helpers without the need of effortfull data collection, robot programming, or domain adaptations. Finally, using mental models for team interactions will significantly improve team performance as well as robot acceptance among human teammates.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 23, 2019
- Source ID
- N000141912311
Entities
People
- Luis Sentis
Organizations
- Office of Naval Research
- United States Navy
- University of Texas at Austin