Learning Goal-Oriented Hierarchical Tasks from Situated Interactive Instruction

Shiwali Mohan, John E. LairdAAAI 2014 · DOI

Introduction

As robots move from controlled industrial settings into dynamic human environments like homes and offices, the ability to learn new tasks from non-expert users becomes critical. Traditional programming is insufficient for the sheer variety of domestic chores, and standard machine learning approaches often require prohibitive amounts of data. This paper presents a framework for "taskable" robots that learn complex, hierarchical tasks through a combination of Explanation-Based Learning (EBL) and real-time dialogue with a human instructor.

The Challenge of Taskable Robots

Most current robotic systems are specialized for specific, pre-programmed actions. Learning from Demonstration (LfD) has emerged as a popular alternative for teaching low-level motor skills, but it struggles with high-level logic and abstract reasoning. LfD is typically data-hungry, requiring numerous examples to discern the underlying structure of a task, and the resulting knowledge is often too specific to transfer to new contexts. The authors address this by focusing on goal-oriented hierarchical tasks—tasks involving multiple sub-goals and conditional logic, such as "to store an object, the pantry must first be open."

The Rosie Framework: Interactive Explanation-Based Learning

The researchers implemented their approach in an agent named Rosie, built on the Soar cognitive architecture. Rosie bridges the gap between low-level execution and high-level reasoning through Interactive Explanation-Based Learning (EBL).

Rosie's situated instruction loop: combining physical action, internal semantic memory, and natural language dialogue.

Rosie represents tasks as hierarchical structures containing parameters, availability conditions, child tasks (sub-steps), and termination conditions. The learning process is situated and interactive:

  1. Immediate Explanation: When faced with a goal, Rosie performs an internal look-ahead search using existing knowledge. If a path is found, she "chunks" the successful sequence into a general rule.
  2. Retrospective Explanation: If internal search fails, Rosie asks the human instructor for the next step. After completing the task, she analyzes her episodic memory to understand why the instructor's suggestions worked, converting specific instances into reusable task knowledge.
  3. Variablization: A critical component of this learning is the replacement of specific objects (e.g., "red block") with variables. This allows Rosie to generalize the logic of a task like "store" to any object and any valid location.

Experimental Evaluation

The framework was evaluated in a simulated robotic environment and validated on a physical robotic arm. The workspace included foam blocks and specific locations like a "pantry," "stove," and "table." Rosie was tested on seven domestic tasks of increasing complexity: place, move, discard, store, cook, stack, and serve.

Key Results and Insights

The evaluation demonstrated that Rosie leverages existing knowledge to learn significantly faster than pure data-driven methods.

Learning efficiency: The number of interactions required to learn the 'cook' task drops to zero as the agent generalizes through variablization.
  • Rapid Generalization: In the cook task, the number of human interactions required dropped from approximately 16 in the first episode to zero by the fifth, as shown in the figure above. The use of variablization was key to this rapid decline.
  • Transfer Learning: Knowledge is cumulative. Learning a basic task like place significantly reduced the interactions needed for complex tasks like stack or serve. For instance, serve required nearly 40 interactions without prior knowledge, but fewer than 10 with transfer.
  • Autonomy through Exploration: As Rosie's internal search depth increased, her self-reliance grew. At a search depth of 4, she could solve the store task entirely through autonomous exploration without instructor intervention.

Conclusion and Future Directions

By combining symbolic reasoning with interactive instruction, the Rosie agent demonstrates a path toward robots that can be taught by non-experts in a single afternoon rather than programmed over months. While the current system is limited to achievement-oriented goals and assumes a perfect teacher, it provides a robust foundation for hierarchical task learning. Future work will focus on handling maintenance tasks, complex logical goals (e.g., disjunctive goals), and more nuanced object categorization.

Summary updated

BibTeX
@inproceedings{mohan2014learning,
  author = {Shiwali Mohan and John E. Laird},
  title = {{Learning Goal-Oriented Hierarchical Tasks from Situated Interactive Instruction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2014},
  doi = {10.1609/aaai.v28i1.8756},
}
Figure