In industrial IoT settings, autonomous mobile robots are deployed for material handling, assembly, and infrastructure inspection. Collaborative perception lets robots share LiDAR observations to build a richer environmental model than any single robot could achieve. Yet robot positions shift, network bandwidth fluctuates, and each robot's contribution to perception quality varies at runtime. Existing methods assume static participation and cannot adapt without sacrificing detection precision or communication efficiency. CoAdapt introduces an adaptive collaborative perception framework where a Large Language Model (LLM) acts as a runtime fusion controller. The LLM decides which robots join the fusion and which fusion algorithm to use based on the current spatial configuration and network state. To feed the LLM, raw LiDAR point clouds are transformed into structured natural‑language descriptions, enabling reasoning without task‑specific training and generalizing to unseen swarm topologies. Evaluation on the OPV2V benchmark across 25 scenarios shows that CoAdapt reduces communication cost by 38% while keeping detection precision comparable to static baselines.
Review