Eluna Automates Warehouse Operations with Agentic LLM System

Warehouse operations are typically governed by intricate Standard Operating Procedures (SOPs) that involve complex decision logic and interactions with multiple systems, all under strict time constraints. Traditional LLM agents often struggle with enforcing procedural compliance and can degrade under the context overload introduced by full SOP specifications. Researchers have developed Eluna, a production-deployed agentic system specifically designed for reliable SOP execution in warehouses. Eluna operates as a graph-guided, multi-agent framework where SOPs are encoded as directed acyclic graphs, allowing for progressive disclosure of information. Independent tasks are delegated to parallel sub-agents, each equipped with persistent code execution and live data access. To meet stringent production latency and accuracy requirements, Eluna employs asymmetric episodic distillation. A powerful teacher model is improved through error memories, and then a smaller student model is fine-tuned on these corrected trajectories, internalizing the improvements without incurring inference-time overhead. Benchmarks across 13 tasks and two production applications show Eluna's fine-tuned models matching or exceeding their teacher, outperforming larger off-the-shelf baselines, and achieving 94% expert agreement in ticket processing.