Skill Learning for Goal-Directed Behavior by Humanoid Robots in Complex Unstructured Environments
Abstract
*Approved for Public Release*The aim of this proposal is to test the hypothesis that learned skillswith abstract symbols can enablea legged robot to 1) learn skillsfrom VR demonstrations 2) carry out complex tasks autonomously such asnavigating an indoor environment and shipboard maintenance and 3)interact with a human teammate using language. A key part of theproblem is enabling robots to manipulate objects in their environmentto carry out complex tasks such as opening doors and drawers,manipulating tools to cut and strip a wire, or riding an elevator.These tasks are challenging because the robot must form a mappingbetween high-dimensional sensor data and low-dimensional task-levelgoals, and then create high-dimensional signals to send to its sensorsand actuators based on thosegoals. It must also do so with theguidance, and under the command of, a human partner; it must thereforealso coordinate with that partner to enable fluid collaboration.This project will focus on the notional problems of patrolling andshipboard maintenance. Patrolling requires a robot to navigate acomplex, unstructured environment designed for humans rather than forrobots. To do so, the robot must be able to open doors, operateelevators, and recognize unusual events and situations. Shipboardmaintenance requires navigating tight spaces, opening drawers,hatches, and using tools such as a screwdriver to unscrew a screw.This project is therefore highlyrelevant to naval missions supportingsailors and Marines in complex environments.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 05, 2021
- Source ID
- N000142112584
Entities
People
- Stefanie Tellex
Organizations
- Brown University
- Office of Naval Research
- United States Navy