Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TwoStep: Multi-agent Task Planning using Classical Planners and Large Language Models

David Bai, Ishika Singh, David Traum, Jesse Thomason

TL;DR

The paper introduces TwoStep, a framework that converts a multi-agent planning problem into parallel single-agent planning tasks by using LLMs to generate helper subgoals and translating them into PDDL. This hybrid approach leverages the strengths of classical planners for execution guarantees while harnessing LLM commonsense to partition goals and enable concurrency. Empirical results across five symbolic domains and an embodied AlfWorld/AI2THOR setup show faster planning times and shorter execution lengths than traditional multi-agent PDDL planning in most cases, with subgoals that closely match human expert specifications. The work demonstrates practical gains in multi-agent coordination and provides prompts for subgoal generation and translation to facilitate future development.

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, such as concurrent actions between two agents when there are no conflicting conditions, without significant modification and definition to existing PDDL domains. A human expert aware of such constraints can decompose a goal into subgoals, each reachable through single agent planning, to take advantage of simultaneous actions. In contrast to classical planning, large language models (LLMs) directly used for inferring plan steps rarely guarantee execution success, but are capable of leveraging commonsense reasoning to assemble action sequences. We combine the strengths of both classical planning and LLMs by approximating human intuitions for multi-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, as well as most multiagent plans, while guaranteeing execution success. Additionally, we find that LLM-based approximations of subgoals result in similar multi-agent execution lengths to those specified by human experts. Website and resources at https://glamor-usc.github.io/twostep

TwoStep: Multi-agent Task Planning using Classical Planners and Large Language Models

TL;DR

The paper introduces TwoStep, a framework that converts a multi-agent planning problem into parallel single-agent planning tasks by using LLMs to generate helper subgoals and translating them into PDDL. This hybrid approach leverages the strengths of classical planners for execution guarantees while harnessing LLM commonsense to partition goals and enable concurrency. Empirical results across five symbolic domains and an embodied AlfWorld/AI2THOR setup show faster planning times and shorter execution lengths than traditional multi-agent PDDL planning in most cases, with subgoals that closely match human expert specifications. The work demonstrates practical gains in multi-agent coordination and provides prompts for subgoal generation and translation to facilitate future development.

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, such as concurrent actions between two agents when there are no conflicting conditions, without significant modification and definition to existing PDDL domains. A human expert aware of such constraints can decompose a goal into subgoals, each reachable through single agent planning, to take advantage of simultaneous actions. In contrast to classical planning, large language models (LLMs) directly used for inferring plan steps rarely guarantee execution success, but are capable of leveraging commonsense reasoning to assemble action sequences. We combine the strengths of both classical planning and LLMs by approximating human intuitions for multi-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, as well as most multiagent plans, while guaranteeing execution success. Additionally, we find that LLM-based approximations of subgoals result in similar multi-agent execution lengths to those specified by human experts. Website and resources at https://glamor-usc.github.io/twostep
Paper Structure (21 sections, 1 equation, 3 figures, 2 tables, 1 algorithm)

This paper contains 21 sections, 1 equation, 3 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background
  3. Classical planning problem
  4. Planning Domain Definition Language (PDDL)
  5. Multi-agent planning problem
  6. Multi-agent Planning Method: TwoStep
  7. English subgoal generation for helper agents
  8. Translation from English description of subgoals to PDDL goal specifications
  9. Editing initial state for main agent
  10. Multi-agent plan execution
  11. Experiment setup
  12. Domains
  13. Symbolic Domains
  14. Embodied Domain
  15. Approaches Considered
  16. ...and 6 more sections

Figures (3)

  • Figure 1: PDDL overview. PDDL planning requires a domain definition, with state predicates (a) and action operators (b), and an initial and goal condition specifications (c). A solver can then use these definitions to generate a plan (d).
  • Figure 2: Planning pipeline for TwoStep. TwoStep has $N$helper agents and a main agent. Each helper agent tries to extract and complete a partially independent subgoal that reduces the steps for main agent, while main agent completes the remaining task after all helpers have completed subgoals. The plans generated for all agents are then executed to find the execution length.
  • Figure 3: Results comparing vanilla planning from SA PDDL ($N=1$) and MA PDDL ($N\geq1$) to TwoStep. The shaded background is the standard deviation across 3 runs for TwoStep, and the rightmost graphs represent the average of the normalized per domain results across all domains. All averages are across all 20 tasks, with the exception of MA PDDL in the TERMES domain, where planning times without producing a plan out for 5/20 tasks with 4 agents. On average, TwoStep reduces planning time and improves parallelization over vanilla multi-agent PDDL planning as the number of agents increases.