AMRF: A Software Architecture for Autonomous Mobile Robot Fleet Management in Warehouse Environments

Coordinating autonomous mobile robots in single-platform warehouses requires tight integration between motion control, fleet orchestration, and warehouse data — yet existing open frameworks (Open-RMF, VDA 5050) are designed for multi-vendor interoperability and impose substantial integration overhead. We present AMRF, a three-layer software architecture that targets the single-platform case: a ROS 2 robot stack, a GraphQL fleet gateway, and a warehouse control system that co-locates A* path planning inside PostgreSQL (via pgRouting) with combinatorial route optimisation as a separate service. Hardware constraints such as lift height and payload are encoded as graph edge attributes, and the entire stack deploys from a single Docker Compose manifest. In a pilot deployment on one AMR in a 15 × 20 m laboratory warehouse, median end-to-end command latency was 16 ms (P95 23 ms), QR+SLAM localisation kept 90% of pre-docking poses inside a 30 mm tolerance with image-based visual servoing closing the remainder (30/30 successful engagements), and the combinatorial dispatcher sustained near-linear throughput scaling across batch sizes 4, 8, and 16. Against batched-FIFO and nearest-neighbour baselines on identical task sets, the OR-Tools VRPPD dispatcher cut total route distance by 48–74% relative to nearest-neighbour and 58–80% relative to batched FIFO, isolating the gain of combinatorial optimisation from batching alone; a direct external-reference measurement put the final docking residual at a mean of 6.6 mm (P90 11.9 mm) over 30 trials; and A* graph-search latency stayed sub-millisecond as the warehouse graph scaled from 62 to 500 nodes. We report the study as a single-robot engineering pilot that validates the integration pattern at single-robot scale; multi-robot fleet validation remains the principal open item.