Source-linked AI summary
Scalable Multi-Robot Collaboration with Large Language Models: Centralized or Decentralized Systems?
Yongchao Chen, Jacob Arkin, Yang Zhang, Nicholas Roy, Chuchu Fan
TL;DR
Multi-robot LLM planning must scale across heterogeneous agents and long horizons without exhausting context budgets. This paper compares four communication frameworks and finds that hybrid approaches achieve stronger success and scalability, with demonstrations extending to 3D coordination.
Problem
Many-robot, long-horizon planning makes coordination harder while increasing context length, inference runtime, API costs, and context dilution.
Method
The paper compares four LLM communication frameworks for incrementally planning actions across increasing numbers of heterogeneous robots.
Results
Hybrid frameworks produce the most successful plans and scale best to large numbers of agents across the studied warehouse-related tasks.
Takeaways & Limitations
Hybrid central-local planning preserves LLM-based task-planning generality while improving multi-robot scalability within the studied scope.
Abstract
from arXiv · showhide
A flurry of recent work has demonstrated that pre-trained large language models (LLMs) can be effective task planners for a variety of single-robot tasks. The planning performance of LLMs is significantly improved via prompting techniques, such as in-context learning or re-prompting with state feedback, placing new importance on the token budget for the context window. An under-explored but natural next direction is to investigate LLMs as multi-robot task planners. However, long-horizon, heterogeneous multi-robot planning introduces new challenges of coordination while also pushing up against the limits of context window length. It is therefore critical to find token-efficient LLM planning frameworks that are also able to reason about the complexities of multi-robot coordination. In this work, we compare the task success rate and token efficiency of four multi-agent communication frameworks (centralized, decentralized, and two hybrid) as applied to four coordination-dependent multi-agent 2D task scenarios for increasing numbers of agents. We find that a hybrid framework achieves better task success rates across all four tasks and scales better to more agents. We further demonstrate the hybrid frameworks in 3D simulations where the vision-to-text problem and dynamical errors are considered. See our project website https://yongchao98.github.io/MIT-REALM-Multi-Robot/ for prompts, videos, and code.
I. INTRODUCTION
Multi-robot planning must handle heterogeneous capabilities, coordination, task dependencies, safety, and increasing difficulty as robot numbers grow. The paper studies LLM dialogue frameworks intended to retain generalization while improving scalability and planning success.
- Multi-robot planning is difficult because robots may have heterogeneous capabilities, coordination requirements, task interdependencies, and safety constraints.
- 6 prior LLM-based efforts assign one LLM to each robot and evaluate collaborative dialogue for systems of only two or three robots.
- Many-robot, long-horizon planning increases coordination complexity, context-token requirements, inference runtime, API costs, and context dilution.
- The study compares four cooperative LLM dialogue frameworks as robot numbers increase, using incremental collaboration to select each agent’s next action.
- The proposed HMAS-2 framework and state-action history method improve planning success rates in situations with high robot numbers.
II. PROBLEM DESCRIPTION
The system converts goals and observations into text for LLM-based planning, executes the resulting robot actions, and feeds the new state into the next planning iteration. Plans are checked for syntax before execution.
- The system assumes N robots and M LLM agents with full environment knowledge and heterogeneous robot capabilities.
- HMAS-1 and HMAS-2 are illustrated through local-agent and central-agent prompt examples, respectively.
- Goals and current environment state are given as text, after which agents dialogue to generate an initial action set.
- Before execution, an external rules-based verifier checks action-set syntax and returns errors as feedback for correction.
- After the robots execute a syntactically correct action set, the resulting environment state becomes context for the next planning iteration.
A. Main Components of LLM Prompt
Each LLM prompt combines task requirements, step history, current state, robot capabilities, role-specific instructions, communication rules, and optional syntax-feedback information.
- The prompt structure includes task description, step history, current state, robot state and capability, agent specialization, communication instruction, and optional syntax feedback.
- Step history records prior dialogue, environment states, and actions from the iterative planning process.
- Robot State & Capability specifies each robot’s available actions and current location, while safe planning remains the planner’s responsibility.
- Agent-specialized prompts emphasize local state and other agents’ responses, while central-agent prompts include local-agent feedback.
- Optional syntactic-checking feedback explains formatting errors and whether generated outputs use available actions.
B. Communication Frameworks for Sub-task Plan
The paper compares decentralized, centralized, and two hybrid communication frameworks for incrementally generating multi-robot sub-task plans. The hybrid designs combine central proposals or feedback with local-agent dialogue while limiting participation to acting agents.
- DMAS assigns each robot an LLM planner and uses turn-taking dialogue whose concatenated comments lengthen prompts until execution is proposed.
- CMAS uses one central LLM to assign actions to every robot at each planning iteration.
- HMAS-1 has a central planner propose initial actions before robot LLMs continue decentralized dialogue, whereas HMAS-2 is a hybrid variant of CMAS.
- Only agents taking an action participate in HMAS-1 and HMAS-2 dialogue, reducing dialogue duration and prompt-token use.
C. Step History
The study compares three historical-context choices for LLM planners because full dialogue and state histories can exhaust the context budget. State-action history offers the best reported trade-off between task performance and token efficiency.
- The ablation compares no history, state-action pair history, and full history in the planners’ context.
- Full dialogue, environment-state, and action histories rapidly exhaust the LLM context token budget.
- State-action pair history has the best reported trade-off between task performance and token efficiency.
D. Token Length Constraint
The experiments evaluate four warehouse-inspired environments across increasing robot counts, while token limits constrain planner context. Trials fail for several execution, dialogue, syntax, planning, or collision conditions.
- The planners use context limits of 8192 tokens for gpt-4-0613 and 4097 for gpt-3.5-turbo-0613.
- The four environments are tested with increasing robot counts, using 4, 8, 16, and 32 robots for BoxNet1 and BoxNet2, and 4, 6, 8, and 10 for Warehouse and BoxLift.
- A trial fails when dialogue exceeds token or round limits, syntax checking or planning exceeds its limit, or the plan causes a collision.
B. Metrics
The evaluation measures planning quality through success and plan length, and efficiency through tokens and API calls. Steps, tokens, and API calls are normalized over successful plans, with 1.0 representing the best value.
- The evaluation reports average task success rate, steps per plan, tokens per plan, and API calls per plan.
- Steps, tokens, and API calls include only successful plans and are therefore reported as normalized values.
- The best normalized metric value is 1.0, so the best-performing framework for a metric receives 1.0.
C. Results
HMAS-2 provides the strongest planning quality, while CMAS minimizes API calls and tokens and DMAS performs worst on success and communication efficiency. HMAS-2 also scales better to more agents and benefits from state-action history, whereas GPT-3 performs worse than GPT-4.
- HMAS-2 achieves the highest success rates and fewest actions per plan, whereas CMAS uses the fewest API calls and tokens.
- DMAS has the most API calls and tokens and the lowest task success rate because its agents often prolong dialogue without adding new information.
- CMAS scales significantly worse than HMAS-2 as agent numbers increase, while HMAS-2 can check and correct flawed plans such as collision-causing actions.
- HMAS-2 performs much worse without prior-history context than with state-action pair history, while full history performs slightly worse than state-action history.
- GPT-3 performs significantly worse than GPT-4, indicating that LLM-based planner quality depends on the underlying model’s quality and capability.
D. 3D Simulation
The study extends evaluation to 3D robot-arm simulations involving colored boxes, bowls, confined workspaces, and collaborative manipulation. Across ten-run scenarios with three and six agents, both methods achieved perfect success, while HMAS-2 used fewer action steps with six robots.
- 3D Task Setup: The 3D task requires robot arms to move colored boxes into matching bowls while operating within assigned workspaces.Arms can pick and place only boxes within their workspace or on its border.
- HMAS-2 Communication: HMAS-2 communicates collision-risk reports from local agents to a central agent.This communication supports the iterative replanning process used by the LLM-based frameworks.
- Results: Both CMAS and HMAS-2 achieve 100% success rates in the three- and six-agent 3D scenarios.Results are averaged over ten runs for each scenario.
- Results: CMAS uses more action steps than HMAS-2 in the six-robot situation.This agrees with the reported pattern that CMAS performs worse than HMAS-2 on more complex tasks.
V. RELATED WORK
Related work applies LLMs to robotics through action selection, motion-planning integration, and reward specification. Early multi-robot studies use collaborative LLM dialogue but remain limited in robot count and do not examine framework or history scalability.
- LLMs for Robotics: Representative robotics systems use LLMs to select actions from predefined skill primitives and complete tasks step by step using text or code intermediates.Examples include SayCan, Inner Monologue, Code-As-Policy, and ProgGPT.
- LLMs for Robotics: Other approaches connect LLMs with task-and-motion planning or use LLM-generated rewards for independent reward-based planners.Reward formats include real values, temporal logics, and patterns.
- Multi-Robot Planning: Recent multi-robot studies assign one LLM to each robot and use collaborative dialogue, but they address only two or three robots.These studies do not consider scalability of communication frameworks or step histories.
VI. CONCLUSION
The work studies scalable LLM-based planning for long-horizon, heterogeneous multi-robot tasks and proposes collaborative dialogue frameworks. Hybrid approaches combining central and local planners produce the most successful plans and scale best to larger agent populations, while future work should investigate more complex hierarchical organizations.
- VI. CONCLUSION: The work targets long-horizon multi-robot task planning with many robots and heterogeneous capabilities.The experiments focus on warehouse-related tasks.
- VI. CONCLUSION: Hybrid frameworks combining central and local LLM planners produce the most successful plans and scale best to larger numbers of agents.The conclusion identifies this pattern across the paper’s multi-robot planning evaluation.
- VI. CONCLUSION: Future work should explore more complex tasks using hierarchical frameworks that organize agents into specialized robot sub-groups.The paper also identifies multimodal, particularly visual, models as a direction for communication-paradigm research.