Improving Knowledge Extraction from LLMs for Task Learning through Agent Analysis

James R. Kirk, Robert E. Wray, Peter Lindes, John E. LairdAAAI 2024 · DOI

Introduction

Large Language Models (LLMs) possess vast general knowledge, making them attractive for robotic task planning. However, they often struggle with "grounding"—suggesting actions that are physically impossible for a specific robot or involve objects not present in the environment. This paper presents STARS, a framework that enables an embodied agent to mentally simulate LLM suggestions, identify mismatches, and autonomously prompt the LLM for repairs.

The Grounding Challenge in Robotics

The authors identify four critical requirements for LLM-generated instructions to be useful in a physical setting:

  1. Interpretability: The robot's internal parser must be able to process the language.
  2. Situational Grounding: Referenced objects must exist in the robot's current environment.
  3. Embodiment/Affordances: Actions must be physically executable by the specific robot hardware.
  4. Human Alignment: Tasks must adhere to specific user preferences.

Traditional Template-Based Prompting (TBP) frequently fails these criteria, necessitating significant manual intervention from human teachers.

The STARS Framework

The STARS framework (Search Tree, Analyze and Repair, and Selection) integrates the LLM into a broader cognitive architecture.

The STARS framework: Search Tree, Analyze & Repair, and Selection.

Search Tree (ST)

Rather than requesting a single response, the agent uses beam search to generate a variety of high-probability goal descriptions from the LLM.

Analyze and Repair (AR)

This core component involves the agent performing a "mental simulation" of each candidate response. It verifies linguistic interpretability, visual grounding of objects, and physical affordances (e.g., checking if a dish rack is grabbable). If a mismatch is detected, the agent provides targeted diagnostic feedback to the LLM to request a corrected suggestion.

Example of the Analyze and Repair process in a simulated kitchen environment.

Selection and Oversight

A more capable LLM (e.g., GPT-4) selects the most reasonable goal from the vetted pool. If a human is available for oversight (STARS+O), the agent presents these high-quality options for simple confirmation, significantly reducing the human's communication burden.

Experimental Evaluation

STARS was evaluated in the APRIL MAGIC simulator, featuring a mobile manipulator. Testing spanned three scenarios—Tidy Kitchen, Store Groceries, and Organize Office—involving 58 unique objects.

Results

  • Autonomous Performance: STARS achieved 77.5% to 94.4% task completion autonomously, a substantial improvement over the TBP baseline (35% to 66%).
  • Human Efficiency: With human oversight (STARS+O), the system reached 100% success while requiring 52-68% fewer words from the human teacher compared to the baseline.
  • Knowledge Retention: Once a goal was learned, the agent could repeat the task with zero further LLM or human intervention.

Conclusion

The STARS framework shifts the robot's role from a passive recipient of instructions to an active critic. By closing the loop with diagnostic feedback and focusing on goal-state learning rather than rigid action sequences, STARS enables more robust and flexible autonomous task acquisition.

Summary updated

BibTeX
@inproceedings{kirk2024improving,
  author = {James R. Kirk and Robert E. Wray and Peter Lindes and John E. Laird},
  title = {{Improving Knowledge Extraction from LLMs for Task Learning through Agent Analysis}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i16.29799},
}
Figure