Each robot plans with its own LLM: a geometric verifier turns conflicts into structured feedback and replans only the affected robot

Multi-robot navigation planning gains another architecture: instead of one LLM reasoning over the whole team, each robot proposes its own symbolic task plan, and a shared execution-level verifier turns conflicts in the grounded trajectories into structured feedback so only the robots involved replan. Earlier multi-robot LLM frameworks coordinated teams through centralized or decentralized symbolic reasoning and inter-agent communication, but symbolic coordination by itself does not ensure the plans stay physically feasible after geometric grounding. In CARLA two-agent multi-target navigation, centralized and decentralized architectures were compared under the same task conditions: centralized planning incurs lower planning overhead for smaller target sets, while the gap narrows for larger target sets as decentralized planning benefits from its modular structure.

Coordination does not rely on robots negotiating in unrestricted natural language; it relies on an external verifier inspecting the grounded trajectories. For every robot pair the verifier checks whether trajectory separation meets a prescribed threshold, and it can additionally check robot-obstacle constraints using the same grounded trajectories. When the condition is violated, the verifier returns the affected agents, the predicted conflict time and location, and the minimum predicted separation as a structured feedback object; replanning is targeted to the affected agent while feasible plans from other agents may be retained, until all candidate trajectories satisfy the execution-level constraints.

The same planning-verification-execution interface carries over from ground vehicles to aerial robots. The CARLA experiments examine coordination performance on a road-network graph, while the Crazyflie experiments move the same symbolic-to-geometric planning and feasibility pipeline onto quadrotors. In two-Crazyflie simulation, two independently tasked quadrotors coordinate through trajectory-level conflict checking and replanning; on a single Crazyflie in hardware, the LLM generates sparse waypoints that are grounded and checked for feasibility before reaching the low-level controller, and the executed trajectory deviates from ideal waypoint connections due to tracking errors, actuator dynamics, localization uncertainty, and external disturbances.

The interface targets high-level task planning for multi-robot navigation: LLMs propose symbolic plans, a geometric layer grounds them into trajectories, and an execution-level verifier checks predicted conflicts and returns structured feedback. CARLA experiments use two-agent multi-target navigation on a structured road-network graph to study coordination performance as task complexity increases, and Crazyflie experiments study transfer of the same planning-verification-execution interface to aerial robots. The authors propose extending the framework to larger robot teams, richer communication policies, uncertainty-aware verification, and multi-robot hardware deployment.

In the present implementation the verifier operates as an execution-level feasibility checker for predicted trajectory conflicts, and the authors list uncertainty-aware or formally certified safety mechanisms as a natural direction for future work. On hardware, the executed trajectory deviates from ideal waypoint connections due to tracking errors, actuator dynamics, localization uncertainty, and external disturbances, so the hardware experiments evaluate the planning-grounding-feasibility interface under execution-level effects while multi-robot coordination itself is evaluated in the two-Crazyflie simulation. The authors also list larger robot teams, richer communication policies, and multi-robot hardware deployment as ongoing work.