Independent robot planners no longer collide: a geometric verifier returns conflict place and time to the right LLM for replanning
Synopsis
Multi-robot task planning no longer relies on symbolic negotiation alone. Each robot proposes a task plan with its own large language model, and a geometric verifier grounds those plans into trajectories, detects predicted inter-robot conflicts, and returns the affected agents, conflict location and conflict time as structured feedback for targeted replanning. In CARLA two-agent multi-target navigation and two-Crazyflie simulation, centralized planning has lower planning overhead for small target sets while the gap narrows as targets grow; neighbor-list inputs consistently yield graph-valid routes, whereas coordinate inputs are more flexible but more prone to invalid waypoints.
Interpretation
Multi-robot planning is split into two layers: the large language model handles semantic task decomposition, while a shared geometric verifier decides whether the trajectories are actually feasible. Earlier multi-robot language-model frameworks coordinate teams through centralized or decentralized symbolic reasoning and inter-agent communication, but symbolic coordination by itself does not ensure plans stay physically feasible after geometric grounding. In CARLA two-agent multi-target navigation, each agent starts at an assigned node, visits all designated targets in any order and returns to its start, and a trial succeeds only if both agents visit every target and stay collision-free, with a meter-scale inter-agent separation threshold.
When the verifier predicts a conflict, it does not make every robot renegotiate; it returns the conflict information only to the affected planner. Independently generated routes can intersect in space and time or violate shared workspace constraints, and these conflicts are not visible at the symbolic planning level. For every robot pair the verifier checks whether the minimum predicted separation falls below a prescribed threshold, and on violation builds a structured feedback object with the affected agents, predicted conflict time and location, and minimum predicted separation; replanning targets the affected agent while other agents keep feasible plans, until all candidate trajectories satisfy execution-level constraints.
The same planning-verification-execution interface carries from ground vehicles to aerial robots, with grounding and execution layers swapped per platform. The CARLA experiments cover only ground vehicles on a structured road network, so they cannot show whether the interface depends on that road-network setting. Two quadrotors in Crazyflie simulation are given distinct navigation tasks and the same centralized and decentralized architectures generate trajectories satisfying the verifier's predicted separation criterion; on physical hardware a single Crazyflie runs the same symbolic-to-geometric pipeline, with the LLM producing sparse waypoints that are grounded into executable trajectories and checked for feasibility before reaching the low-level controller.
Perspective
The work targets high-level task planning for multi-robot navigation, in settings where a robot team shares a workspace and must avoid spatial and temporal conflicts. The CARLA experiments cover two agents on multi-target navigation with increasing target counts, while the Crazyflie portion evaluates two-robot coordination in simulation and single-robot obstacle-aware navigation on hardware. The authors propose extending the framework to larger robot teams, richer communication policies, uncertainty-aware verification and multi-robot hardware deployment.
The verifier currently acts as an execution-level feasibility checker for predicted trajectory conflicts and does not yet include uncertainty-aware or formally certified safety mechanisms. On hardware, the executed trajectory naturally deviates from ideal waypoint connections because of tracking errors, actuator dynamics, localization uncertainty and external disturbances, so multi-robot coordination itself is evaluated in the two-Crazyflie simulation.
