TwoStep: Multi-agent Task Planning using Classical Planners and Large Language Models (2403.17246v1)

Abstract: Classical planning formulations like the Planning Domain Definition Language (PDDL) admit action sequences guaranteed to achieve a goal state given an initial state if any are possible. However, reasoning problems defined in PDDL do not capture temporal aspects of action taking, for example that two agents in the domain can execute an action simultaneously if postconditions of each do not interfere with preconditions of the other. A human expert can decompose a goal into largely independent constituent parts and assign each agent to one of these subgoals to take advantage of simultaneous actions for faster execution of plan steps, each using only single agent planning. By contrast, LLMs used for directly inferring plan steps do not guarantee execution success, but do leverage commonsense reasoning to assemble action sequences. We combine the strengths of classical planning and LLMs by approximating human intuitions for two-agent planning goal decomposition. We demonstrate that LLM-based goal decomposition leads to faster planning times than solving multi-agent PDDL problems directly while simultaneously achieving fewer plan execution steps than a single agent plan alone and preserving execution success. Additionally, we find that LLM-based approximations of subgoals can achieve similar multi-agent execution steps than those specified by human experts. Website and resources at https://glamor-usc.github.io/twostep

Multi-agent Task Planning using Classical Planners and LLMs

Introduction

In the presented work, a novel approach for facilitating multi-agent planning, named TwoStep, is introduced. This method integrates the reliability of classical planning, as implemented through the Planning Domain Definition Language (PDDL), with the adaptive, commonsense reasoning exemplified by LLMs. The primary innovation in TwoStep is its method for goal decomposition that approximates human intuition in splitting a task among agents. By allowing LLMs to predict subgoals for individual agents which, when executed collectively, address a single, larger goal, TwoStep manages to leverage the benefits of both planning paradigms.

Classical Planning and Its Limitations

The foundational aspect of classical planning has been its deterministic, finite, and fully observable state space which guarantees successful plan formulation under given constraints. Notably, frameworks such as STRIPS and its successors built the groundwork for autonomous task planning in various domains. However, these traditional approaches predominantly focus on single-agent scenarios and heavily rely on human expertise for multi-agent planning domain specification. The expansion to multi-agent systems introduces significant complexity, not least because of the exponential increase in the possible actions and states with the addition of each agent.

Leveraging LLMs in Planning

Recent advancements have shown LLMs' potential in guiding action sequences or acting as surrogate planners. Despite their inability to guarantee execution success, LLMs introduce a novel dimension to planning by incorporating commonsense reasoning and contextual understanding. Their versatility has been explored in various innovative approaches, including translating between natural language goals and PDDL specifications, showing promise in scenarios where dynamic, commonsense adaptation is required.

The TwoStep Method

TwoStep operates on the principle of decomposing a multi-agent planning problem into two correlated single-agent planning problems, utilizing LLMs for intelligent goal decomposition. It comprises a two-step process:

1. Subgoal Generation: An LLM predicts a feasible subgoal for a designated "helper" agent, providing a portion of the larger goal that can be accomplished independently and in parallel to the main agent's tasks.
2. Subgoal Translation and Execution: The LLM translates this subgoal into a PDDL-compatible format, after which classical planners generate and execute plans for both the helper and main agents, ensuring the overall goal is attained efficiently.

Empirical Evaluation

The evaluation spanned multiple symbolic planning domains and a simulated environment, contrasting TwoStep against single-agent PDDL (SA PDDL) and multi-agent PDDL (MA PDDL) planning in terms of planning time and execution efficiency. Across these domains, TwoStep demonstrated superior planning efficiency and, in several cases, outperformed SA PDDL in execution steps, presenting a balanced approach to harnessing LLM's adaptive planning insight and classical planning's execution guarantees.

Strengths, Implications, and Prospects

TwoStep exemplifies a pragmatic fusion between the deterministic robustness of classical planning and the dynamic, context-aware capabilities of LLMs. It highlights the potential of using LLMs for complex decision-making and planning tasks, positioning them not as direct planners but as enhancers to classical planning methodologies through intelligent task decomposition and goal prediction. The implications of this research extend beyond academic interest, offering tangible methodologies that could streamline multi-agent systems' deployment in real-world applications. Future work may explore deeper integrations, scalability issues, and broader domain applicabilities, potentially encompassing non-deterministic planning environments and further leveraging LLMs for dynamic problem-solving and adaptation.

Conclusion

This work contributes significantly to the landscape of automated planning, showcasing how the synergistic application of LLMs and classical planning can address the intricacies of multi-agent systems. By introducing a methodology that mitigates the complexities inherent in multi-agent planning while leveraging the contextual reasoning strengths of LLMs, TwoStep opens new avenues for research and application in AI-driven task planning.