WarehouseAI: Knowledge Graph-Grounded Multi-Agent Reasoning Framework for Simulation-Driven Warehouse Planning

Himabindu Thogaru, Anirudh Deodhar, Purvaja Ammanabrolu

Keywords: multi-agent systems, enterprise AI agents, knowledge graph grounding, discrete event simulation, warehouse planning, closed-loop planning

TL;DR: WarehouseAI is a closed-loop multi-agent system that automatically diagnoses warehouse bottlenecks, generates grounded recommendations, and validates improvements by combining Simulations with Knowledge graphs

Abstract: Enterprise AI agents often struggle to accurately reason over dense, multi-stage relational logs produced by industrial simulators. In this paper, we introduce Warehouse AI, a closed-loop multi-agent system designed for automated warehouse optimization built on a Knowledge Graph-grounded framework. The architecture translates natural language intent into simulation parameters, executes a SimPy-driven Discrete Event Simulation (DES), and maps the resulting event logs into a Neo4j knowledge graph. At its analytical core, a dual-path Knowledge Graph Question Answering (KG-QA) reasoning chain processes both operational queries through direct Cypher retrieval and investigative queries through adaptive multi-hop evidence accumulation, with semantically decomposed query generation and two-level self-correction. Orchestrated via a hierarchical LangChain network, six specialized agents coordinate a complete planning cycle: Configure $\to$ Simulate $\to$ Analyze $\to$ Recommend $\to$ Validate. Across 9 independent trials spanning 3 bottleneck scenarios, each designed with non-obvious root causes where surface symptoms point to the wrong resource tier, the system achieves 100\% orchestration accuracy, simulation validity, recommendation grounding and loop improvement rate, with composite evaluation scores of 0.780-0.859 and package processing time reductions of up to 30\% per cycle. The KG-QA engine achieves Pass@1 = 0.92 on a 25-question benchmark, substantially outperforming single-pass generation (0.59) and self-reflection baselines (0.70). Planning loops demonstrating that structural grounding and hierarchical agent orchestration together deliver the reliability and efficiency required for autonomous industrial decision-making.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 10