Exploring large language models for task planning in an open world

Abstract

This thesis builds on the Common Sense-Based Open World Planning (COWP) framework, which integrates a classical task planner with an LLM module to enable robot autonomy in open world household environments. The framework combines the robustness of traditional planning approaches with the unfolding capabilities of LLMs in order to relax closed-world assumptions and handle failures that arise when a robot must handle new information about its environment. The objectives of this work include formulating a general household domain in PDDL for future learning and skill transfer, implementing a more efficient approach to augment a robot’s world knowledge, extending the framework to handle unknown objects, and evaluating the framework’s robustness.

Experimental results demonstrate an 18% improvement in handling open world sit- uations compared to the original framework. The adapted framework also shows

significant reductions in computation time (4.9 times less than the benchmark) and in the number of API calls (over 13 times fewer) required per task and situation simulated. Key findings highlight a potential for improved efficiency and accuracy in robot task planning in open worlds using LLMs but underscore the need for higher con- sistency, better informed models, and more robust collaboration schemes to achieve practical, real-world applications. Future work could focus on refining knowledge representation systems used in the framework, enhancing the LLM’s search capabili- ties, optimizing the prompting strategy, and incorporating mid- and low-level control in the framework to address the granularity of the problem-solving challenge.