Organizations: Department of Mechanical Science and Engineering, University of Illinois Urbana-Champaign, Urbana, IL, USA · Department of Mechanical Engineering, University of Nevada, Reno, NV, USA
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.
Figures & tables
Fig. 1 : Verifier-mediated multi-robot planning architecture. In the centralized architecture, a single LLM jointly generates symbolic plans for the robot team, whereas in the decentralized architecture each robot is assigned an independent LLM planner. The resulting symbolic plans are grounded into robot trajectories and evaluated by a shared execution-level verifier. Predicted obstacle or inter-agent conflicts are converted into structured feedback and returned to the affected LLM planner for targeted replanning.
Fig. 2 : CARLA multi-agent, multi-target navigation environment. The two agents navigate the shared road-network graph while visiting target sets of increasing complexity, with ∣W∣∈{2,5,9} corresponding to the green, red, and black task configurations, respectively.
Fig. 3 : Performance on two-agent multi-target navigation in CARLA as the number of target points increases from 2 to 5 to 9. Results compare centralized single-LLM planning with coordinate and neighbor-list representations against decentralized two-LLM planning with the neighbor-list representation. Metrics are (a) mean planning time of successful trials, (b) mean successful path length, and (c) mean planning time over all trials.
Fig. 4 : Cross-platform evaluation on Crazyflie aerial robots. (a) Two-Crazyflie coordinated navigation in simulation with obstacle and predicted inter-agent conflict checking (demo video: https://tinyurl.com/cn3anffv ). (b) Physical Crazyflie navigation demonstrating transfer of the symbolic-to-geometric planning and feasibility interface to hardware (demo videos: https://tinyurl.com/476s68br ). Multi-robot coordination is evaluated in simulation, while the hardware experiment evaluates the planning pipeline under tracking, localization, actuation, and physical execution effects.