Extending Interactive Task Learning
Abstract
Our goal is creating artificial agents that can interact with humans and learn completely new tasks through instruction. Solving this problem requires integrating man:' capabilities across AI. In this project, we identified three dimensions of task complexity (diverse types of actions, task formulations, and task modifiers), and implemented extensions that demonstrate greater learning capabilities for each dimension than previous work. First, we extended the representations and learning mechanism for innate tasks so the agent can learn tasks that utilize many different types of actions beyond physical object manipulation, such as communication and mental operations. Second, we implemented a novel goal-graph representation so that an instructor can formulate a task as achieving a goc.l and let the agent use planning to execute it, or can formulate the task as executing a procedure, or sequence of steps, when it is not easy to define a goal. Third, we added support for learning subtasks with various modifying clauses, such as temporal constraints, conditions, or looping structures. Our system has been used with various robotic platforms, and it combines all of these extensions while learning complex hierarchical tasks that cover extended periods of time and demonstrate significant flexibility.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 22, 2021
- Accession Number
- AD1224115
Entities
People
- John E. Laird
Organizations
- University of Michigan