Adaptive Symbiotic Information-Sharing Framework Using LLMs for Heterogeneous Multi-Robot Collaboration

Sammanfattning

Autonomous warehouse systems increasingly rely on heterogeneous teams of agents to coordinate transport and support operations. In this thesis, the focus is a task-assignment warehouse setting with automated guided vehicles (AGVs) and picker agents. AGVs act as carrier agents that move racks between storage and delivery locations, while picker agents provide support for cooperative loading and unloading operations. In this setting, an AGV-side rack decision is only useful when compatible picker-side support is available. This thesis investigates whether allowing picker-side evaluation to influence final AGV-side commitment improves coordination in such a heterogeneous warehouse planning problem. The study evaluates a mutualistic staged large language model (LLM) coordination planner in the TA-RWARE environment. The planner separates coordination into AGV-side proposal generation, picker-side support evaluation, and final commitment formation. It is compared with a non-mutualistic staged LLM baseline that uses the same outer execution infrastructure but fixes the AGV-side rack target before picker-side response. A heuristic warehouse policy is also included as a domain-grounded throughput reference. Experiments are conducted across three AGV--picker ratios: 3:6, 6:6, and 6:3. The results show that the mutualistic planner achieves stronger outcomes than the non-mutualistic baseline in the picker-abundant and balanced settings, with higher completed deliveries and shorter assignment execution time. However, this advantage does not persist under picker scarcity. The thesis therefore supports a bounded conclusion: pre-commitment picker-side evaluation can improve staged LLM coordination when it can lead to useful supported alternatives, but its benefit is limited when picker capacity becomes the main bottleneck.