Dexterous Manipulation Specification via Language and Context Constraints
Abstract
A major challenge for robots acting intelligently and dexterously is the precise specification of complex tasks, such as assembly or selection of parts or tools. Human task specifications are never self-contained because they leverage observed context. Left unstated is which context to include as part of the specification. Depending on interpretation, a simple request may be ambiguously underspecified or it may be overspecified to the point of impossibility. We propose a transformative new approach to posing dexterous manipulation problems via a hierarchical system of constraints and solving them with a constrained task planner. We integrate natural language as a specification modality that implicitly leverages this constraint structure to allow increased or decreased goal specificity as needed during implementation of the specification. We also derive methods for inferring task and motion constraints from analogous previous experience, and integrate constrained task planners for exploring multiple strategies for dexterous manipulation. These capabilities will lead to highly flexible and dexterous manipulation capabilities.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 03, 2016
- Source ID
- N000141612080
Entities
People
- Ross Knepper
Organizations
- Cornell University
- Office of Naval Research
- United States Navy