PaperlayerView exact revision on arXiv

Source-linked AI summary

Verifier-Mediated Multi-LLM Coordination for Multi-Robot Path Planning

Vivek Khatana, Zijian Song, Naira Hovakimyan, Petros Voulgaris

arXiv:2610.12761v1cs.ROeess.SY

TL;DR

Independently generated symbolic plans may conflict after geometric grounding, leaving multi-robot feasibility unresolved by symbolic coordination alone. The paper compares centralized and decentralized LLM planning with verifier-guided replanning, finding task-dependent coordination trade-offs in CARLA and demonstrating the interface on Crazyflie robots.

Abstract

from arXiv · show

Foundation models are increasingly used for high-level multi-robot task planning, yet independently generated symbolic plans can become jointly infeasible after geometric grounding because of shared spatial and temporal constraints. We study execution-aware coordination of large-language-model (LLM) planners for multi-robot navigation. We compare a centralized architecture, in which one LLM jointly reasons over the robot team, with a decentralized multi-LLM architecture in which each robot independently proposes a task plan. A geometric verifier grounds the candidate plans into robot trajectories, detects predicted inter-robot conflicts, and converts conflict information into structured feedback for targeted replanning. This verifier-mediated architecture separates semantic task reasoning from execution-level feasibility while enabling iterative coordination among LLM planners. Experiments in CARLA and on Crazyflie aerial robots evaluate coordination performance as task complexity increases and characterize the trade-offs between centralized reasoning, decentralized verifier-mediated coordination, and physical executability.

I. INTRODUCTION

The introduction frames a gap between symbolic multi-robot plans and physical feasibility after geometric grounding. It proposes verifier-mediated coordination and studies centralized and decentralized LLM planning across CARLA and Crazyflie.

I. INTRODUCTION

  • Centralized planning jointly reasons over the team, while decentralized planners offer modular local reasoning but need a mechanism to resolve conflicts between independent plans.Centralized context and decision space grow with team and task complexity.Source: §I · pp. 1–2
  • The central problem is that individually meaningful symbolic plans can become jointly infeasible after grounding because routes may conflict in space, time, or shared workspace.Prior symbolic coordination does not by itself ensure physical feasibility after geometric grounding,.Source: §I · p. 1
  • The verifier checks grounded trajectories for obstacle and inter-robot conflicts, then returns structured, agent-specific feedback for targeted replanning.The feedback can identify affected agents, conflict location, and conflict time.Source: §I · pp. 1–2, Figure 1 · p. 1
  • The paper studies coordination and task-completion performance as CARLA tasks grow in complexity and demonstrates transfer of the planning–verification–execution framework to Crazyflie aerial robots.It distinguishes its use of an external execution-level verifier and agent-specific geometric feedback from prior symbolic coordination and feedback-guided refinement,, –.Source: §I · p. 2

II. VERIFIER-MEDIATED MULTI-ROBOT PLANNING

The planning representation combines a symbolic environment graph and robot-specific tasks with LLM-generated plans, which are grounded into geometric trajectories by a local planner.

A. Multi-Robot Task and Path Representation

  • The environment is represented as a graph whose nodes are admissible symbolic states and whose edges are valid transitions; each robot receives a task specifying its start, goal, required targets, and constraints.The task is denoted τ_i := {s_i, g_i, W_i, F_i, C_i}.Source: A. Multi-Robot Task and Path… · p. 2
  • The LLM generates a symbolic plan using planner context that may include prior execution-level feedback.Source: A. Multi-Robot Task and Path… · p. 2
  • A grounding map converts symbolic states to geometric anchors, and a local geometric planner connects the grounded waypoints into a robot route.An A⋆-based planner is given as an example.Source: A. Multi-Robot Task and Path… · p. 2

B. Centralized and Decentralized LLM Architectures

The centralized architecture uses one LLM to plan for the team, while the decentralized architecture gives each robot its own planner and couples their candidate trajectories through a shared verifier.

B. Centralized and Decentralized LLM Architectures

  • One LLM jointly generates plans from the team task in the centralized architecture; decentralized planners plan independently and couple candidate trajectories through a shared verifier.Joint planning exposes the centralized LLM to the global coordination problem but increases the size of its planning context.Source: B. Centralized and Decentralized LLM… · p. 2
  • The CARLA environment has two agents navigating a shared road-network graph among target sets of 2, 5, or 9 points.The configurations are shown in green, red, and black, respectively.Source: Figure 2 · p. 2

C. Execution-Level Conflict Verification and Feedback

The verifier checks grounded robot trajectories against execution-level constraints and feeds predicted conflicts back for targeted replanning.

C. Execution-Level Conflict Verification and Feedback

  • The verifier evaluates every robot pair’s trajectories against a prescribed minimum-separation threshold and can also check robot-obstacle constraints.A violation prompts the verifier to identify a predicted conflict and construct structured feedback.Source: C. Execution-Level Conflict… · p. 2
  • Feedback includes the predicted conflict time and location, plus the minimum predicted separation between the affected robots.The superscript r denotes the planning iteration.Source: C. Execution-Level Conflict… · p. 2
  • Replanning targets the affected agent while feasible plans from other agents may be retained, and ends when all trajectories satisfy execution-level constraints.Source: C. Execution-Level Conflict… · p. 2

A. CARLA Experiments

The CARLA evaluation compares centralized and decentralized planning for two-agent navigation as target-set size increases, measuring planning time and successful path length. It reports trade-offs in overhead and route validity across architectures and environment representations.

A. CARLA Experiments

  • The CARLA task assigns each of two agents a start, targets to visit in any order, and a return to its start on a shared road network; trials require collision-free completion.Target-set sizes are 2, 5, and 9; distances below 7 m count as unsuccessful.Source: A. CARLA Experiments · pp. 2–3
  • The comparison uses centralized single-LLM planning with coordinate or neighbor-list inputs and decentralized two-LLM planning with neighbor-list inputs.Decentralized agents exchange tentative paths; detected conflict time and location are sent to the corresponding LLM for replanning while the other agent retains its plan.Source: Figure 3 · p. 3, A. CARLA Experiments · p. 3
  • The three metrics are mean planning time over successful trials, mean successful path length, and mean planning time over all trials.For decentralized planning, planning time is averaged across the two LLM planners.Source: A. CARLA Experiments · p. 3
  • For smaller target sets, centralized planning has lower planning overhead, while the gap narrows as target sets grow; neighbor-list inputs yield graph-valid routes with path lengths comparable to centralized planning.Fig. 3 compares planning time and successful path length across target-set sizes and planning representations.Source: A. CARLA Experiments · p. 3, Figure 3 · p. 3

B. Insights from CARLA Experiments

The CARLA experiments identify trade-offs in planning overhead, environment representation, and task complexity. More targets make collision-free completion and conflict resolution harder.

B. Insights from CARLA Experiments

  • Centralized planning is computationally efficient for the two-agent tasks considered, but its performance gap over decentralized planning becomes smaller as task complexity increases.The paper suggests decentralized planning’s modular structure partially offsets its additional coordination overhead as complexity rises.Source: B. Insights from CARLA Experiments · p. 3
  • Neighbor-list inputs consistently produce graph-valid feasible routes with similar path lengths across architectures, while coordinate inputs can yield shorter trajectories but more invalid or out-of-map waypoints.This is a trade-off between geometric flexibility and symbolic validity.Source: B. Insights from CARLA Experiments · p. 3
  • As target points increase, longer trajectories, larger waypoint-ordering spaces, and more frequent inter-agent conflicts make collision-free completion and iterative conflict resolution harder.The findings highlight the importance of explicit execution-level feasibility checking.Source: B. Insights from CARLA Experiments · p. 3

C. Crazyflie Cross-Platform Evaluation

The Crazyflie evaluation tests two-robot coordination in simulation and transfer of the planning–verification–execution interface to aerial hardware. The hardware trial evaluates the pipeline under physical execution effects, not multi-robot coordination.

C. Crazyflie Cross-Platform Evaluation

  • Two Crazyflies coordinate in simulation through synchronized-distance checking and verifier-triggered replanning, while hardware testing deploys the planning pipeline for obstacle-aware navigation.The physical trial uses one Crazyflie and evaluates feasibility under tracking errors, actuator dynamics, localization uncertainty, and external disturbances.Source: C. Crazyflie Cross-Platform Evaluation · pp. 3–4
  • Fig. 4 contrasts two-robot coordination in simulation with physical Crazyflie navigation that transfers the symbolic-to-geometric planning and feasibility interface to hardware.The figure caption distinguishes simulation-based multi-robot coordination from hardware evaluation under tracking, localization, actuation, and physical execution effects.Source: Figure 4 · p. 4, C. Crazyflie Cross-Platform Evaluation · p. 4

D. Insights from Crazyflie Experiments

Crazyflie experiments show verifier-mediated coordination in simulation and transfer of the planning interface to real aerial hardware, despite execution-level effects.

D. Insights from Crazyflie Experiments

  • Two independently tasked quadrotors coordinate in simulation through trajectory-level conflict checking and replanning, extending the interface beyond CARLA’s road-network setting.Physical experiments show the same symbolic-to-geometric planning and feasibility pipeline deployed on a single aerial robot despite tracking errors, actuator dynamics, localization uncertainty, and external disturbances.Source: D. Insights from Crazyflie Experiments · p. 4
  • The results support retaining the high-level planning architecture across robot embodiments while adapting geometric grounding and execution layers to each platform.The verifier currently checks predicted trajectory conflicts; uncertainty-aware or formally certified safety mechanisms are future-work directions.Source: D. Insights from Crazyflie Experiments · p. 4

IV. CONCLUSION

The conclusion presents verifier-mediated coordination as an interface between foundation-model planning and execution-level feasibility. CARLA studies coordination as task complexity increases, while Crazyflie experiments demonstrate transfer to aerial navigation.

IV. CONCLUSION

  • Centralized or decentralized LLM planners propose robot-level plans, and a geometric verifier returns structured conflict information for targeted replanning.This separates high-level semantic reasoning from execution-level feasibility.Source: §IV · p. 4
  • CARLA experiments study coordination under increasing task complexity, while Crazyflie experiments demonstrate transfer of the planning–verification–execution interface to aerial navigation.Source: §IV · p. 4
  • Future work includes larger robot teams, richer communication policies, uncertainty-aware verification, and multi-robot hardware deployment.Source: §IV · p. 4
Loading 2610.12761v1…