AssetOpsBench

AI Agents for Industrial Asset Operations & Maintenance

A unified, open framework for building, orchestrating, and evaluating domain-specific AI agents in Industry 4.0.

📄 Paper · 🤗 Dataset · 🎮 Playground · 📢 IBM Blog · 🎥 Video · 📊 Kaggle · 🚀 Colab

[!IMPORTANT] 🎉 AssetOpsBench is officially accepted at KDD 2026 (Datasets & Benchmarks Track), Jeju, South Korea, alongside our hands-on tutorial Building Reliable Industrial Agents with MCP. See Publications for the full list of 2025–2026 work.

---

At a Glance

Built for: maintenance engineers, reliability specialists, facility planners, and Industry 4.0 researchers. Powered by: LLMs + Time Series Foundation Models, orchestrated over live sensor data and Industry 4.0 records (FMEA, work orders, alerts). Now with: simplified interface and native MCP (Model Context Protocol) support.

---

Quick Start

# Clone and install
git clone https://github.com/IBM/AssetOpsBench.git
cd AssetOpsBench
pip install -e .

# Try a scenario (to be enabled)
python -m assetopsbench.run --scenario "List all sensors of Chiller 6 in MAIN site"

Or jump in instantly:

• 🚀 Run on Colab — no install required (illustration of LLM Agent)
• 🎮 Try the HF Playground — interactive demo
• 📖 Read INSTRUCTIONS.md — full setup, MCP servers, plan-execute runner

[!NOTE] Active development is on main. The codebase used for various publication venues continues to be maintained on separate branches, for example, ACL 2026 IndustryAssetEQA and prior experimental work is maintained on main-0.x.

---

What is AssetOpsBench?

AssetOpsBench is a unified framework for developing, orchestrating, and evaluating domain-specific AI agents in industrial asset operations and maintenance. It provides reproducible scenarios, agent tooling, and evaluation pipelines for multi-step workflows in simulated industrial environments.

Domain-Specific MCP Servers

Five domain servers (IoT, FMSR, TSFM, WO, Vibration) plus one shared utility server, documented in full — arguments, categories, and backing services — in docs/mcp-servers.md.

The full TSFM MCP surface currently contains 41 tools covering model cards, feature cards, recipe execution, evaluation, and result/run lookup. See docs/mcp-servers.md for the complete reference.

The WO MCP surface contains 15 tools (9 read + 6 write) covering the full Maximo-style work-order lifecycle — query, create, approve, assign, close, cancel — plus costs, KPIs, and scheduling, backed by CouchDB with IBM Maximo mxwo field names. Set AOB_READONLY=1 to expose the 9 read tools only. See docs/mcp-servers.md for the complete reference.

Agent Frameworks

• Plan Execute — plan-and-execute sequential workflow to work with any LLM
• Deep Agent — planning, sub-agents, and virtual filesystem for long-horizon tasks
• Claude Agent — ReAct-based orchestrator using Claude with agent-as-tool delegation
• OpenAI Agent — ReAct-based orchestrator using OpenAI models with agent-as-tool delegation

MCP Environment

The src/ directory contains MCP servers and a plan-execute runner built on the Model Context Protocol. See INSTRUCTIONS.md for setup.

---

Example Scenarios

Some tasks focus on a single domain, others are multi-step end-to-end workflows. Explore all scenarios on Hugging Face.

---

Leaderboards

• To be revised (WIP with latest models)
• Evaluated with 7 Large Language Models
• Trajectories scored using LLM Judge (Llama-4-Maverick-17B)
• 6-dimensional criteria measuring reasoning, execution, and data handling

Example: MetaAgent leaderboard

---

Publications

12+ contributions across 7 top venues in 2025–2026 from the team behind AssetOpsBench.

---

Tutorials & Technical Material

📘 Hands-on guides from our team:

• ReActXen IoT Agent (EMNLP 2025)
• FailureSensorIQ (NeurIPS 2025)
• AssetOpsBench Lab (AAAI 2026)
• SPIRAL (AAAI 2026)
• AssetOpsBench Technical Material

---

AI Competitions

AssetOpsBench powers public AI agent competitions that bring together researchers, students, and practitioners worldwide.

🔴 Live — IJCAI 2026

Industrial Automation Challenge: Benchmarking Physics-Grounded LLMs for Task Reasoning

A new challenge co-located with IJCAI 2026 that pushes LLM agents on physics-grounded industrial reasoning.

• 🌐 Challenge site: ai-industrial-challenge-ijcai
• 📋 IJCAI 2026 competitions: 2026.ijcai.org/competitions

✅ Completed — CODS 2025

AssetOpsBench-Live: AI Agentic Challenge

Launched in September 2025 at CODS 2025, the competition evaluated multi-agent systems on live industrial scenarios.

• 🏆 Competition page: codabench.org/competitions/10206
• 👥 365 participants · 500+ agent submissions

---

Talks & Events

---

University Projects & Extensions

AssetOpsBench is being extended by university research groups exploring new asset classes, evaluation paradigms, and agentic architectures. To list your project, open a PR.

• Calibrated Coordination Reduces Overconfident Errors in Multi-Agent LLM Systems – Confidence-weighted aggregation and abstention framework for reducing hallucinated confidence events in multi-agent industrial troubleshooting and operational decision-making benchmarks. Chand Sahil Mansuri, Sadamori Kojaku, Binghamton University.
• Internalizing MCP Tool Knowledge in Small LLMs via QLoRA Fine-Tuning — HPML project using AssetOpsBench to fine-tune ~4B models to internalize MCP tool knowledge and reduce prompt schema overhead. Ayal Yakobe, Columbia University · repo
• SPIN — Structural LLM Planning via Iterative Navigation for Industrial Tasks. Yusuke Ozaki, University at Albany · paper · repo
• Synthetic Scenario Generation for Evaluation of Industry 4.0 Agents — Automated scenario generation, transformer asset integration, and scenario quality evaluation. Rohith Kanathur, Sagar Chethan Kumar, Columbia University · repo · paper
• AgentOpsBench — High-throughput battery analytics MCP server with DNN prognostics (RUL prediction) and 3.3× latency optimization. Siddharth Gowda, Rushin Bhatt, Aryaman Agrawal, Winston Li, Columbia University · repo
• Skill-Knowledge-Augmented Agents on AssetOpsBench — Confidence-gated skill execution with scoped knowledge plugins for industrial fault diagnosis. Vera Mazeeva, Sanskruti Shejwal, Shrey Arora, Mana Abbaszadeh, Columbia University · repo
• Evaluating Temporal Semantic Caching and Workflow Optimization in Agentic Plan-Execute Pipelines. Krish Veera, Alimurtaza Mustafa Merchant, Sajal Kumar Goyla, Shambhawi Bhure, Columbia University · paper · repo
• Towards Multi-Turn Dialog Systems for Industrial Asset Operations and Maintenance - Improved response quality and reduced redundant tool calls and multi-turn latency. Chengrui Li, Rujing Li, Yitong Bai, Rui Li, Columbia University ·paper· repo
• Skills and Knowledge Plugin MCP Servers for Optimized Industrial O&M Agents - reducing planning overhead and improving retrieval grounding in industrial asset maintenance agents through an MCP Skills Server that exposes reusable multi-step operational workflows and a Knowledge Plugin Server that enables injection of context-specific documentation. Andrew Li, Kirthana Natarajan, Thai On, Trisha Maturi, Yeshitha Bhuvanesh, Columbia University · repo
• Profiling and Optimizing the TSFM MCP Server - Developed a reproducible benchmarking harness, stage-level profiling system, and interchangeable model interface that identified preprocessing and inference bottlenecks, achieving up to 12.8× faster forecasting and 12.2% lower fine-tuning latency while supporting forecasting, fine-tuning, and anomaly detection workflows. Tomas Pasiecznik, Sam Colman, Byeolah Kwon, Sally Go, Columbia University · repo
• Profiling and Optimizing the AssetOpsBench Plan-Execute Pipeline - Provides the first systematic performance characterization of the AssetOpsBench plan-execute pipeline to quantify the latency-accuracy tradeoff of thinking mode on Gemma 4 26B for industrial asset operations tasks. Implemented and evaluated scenario-based routing optimizations to balance the tradeoff. Shen Li, Charles Xu, Ann Li, Caroline Cahill, Columbia University · repo
• Performance Optimzation of the TSFM Agent in an Industrial Agentic Benchmark - Developed an optimization framework for IBM’s TinyTimeMixer(TTM) model by implementing model pre-loading, torch.compile graph fusion, and replacing Huggingface abstractions with direct batched model calls. We achieved 3.3X reduction in workflow latency and 68% decrease in total execution time while maintaining zero-shot forecast quality on industrial sensor data. Alisha Vinod, Jonathan Ang, Sanjaii Vijayakumar, Thomas Ajai, Columbia University . repo
• Visual Inspection Agent for AssetOpsBench - Adds a vision modality to AssetOpsBench via an MCP-connected Visual Inspection Agent and 22 hand-authored visual inspection scenarios across pumps, induction motors, power transformers, and wind turbine blades. Benchmarks AWQ W4A16 quantization and vLLM serving optimizations on Qwen2.5-VL-7B and Llama-3-LLaVA-NeXT-8B, with an LLM-as-a-judge scoring pipeline for accuracy evaluation. Amaan Sheikh, Aman Upganlawar, Madhav Rajkondawar, Yang-Jung (Eric) Chen, Columbia University · repo
• Agentic AI Workflows for Naval Operations and Maintenance — Exploring AssetOpsBench for evaluating agentic AI workflows, with future extensions using digital-twin-generated synthetic data. Priyam Dalmia, Chin-Teng Lin, Fred Chang, University of Technology Sydney
• Knowledge Graphs as the Missing Data Layer for LLM-Based Industrial Asset Operations — Treats a typed knowledge graph as the data layer for AssetOpsBench agents: holding the model and orchestration fixed, the same GPT-4 rises from 65% to 82–83% with LLM-generated Cypher and to 99% with deterministic graph handlers on the 139-scenario snapshot, plus Generation-Augmented Knowledge (GAK) — provenance-tagged enrichment — for the non-deterministic FMSR scenarios. Accepted at Agents+Graph @ VLDB 2026. Madhulatha Mandarapu, Sandeep Kunkunuru (VaidhyaMegha / Samyama) · paper · repo

---

Call for Scenario Contribution

We are expanding AssetOpsBench to cover a broader range of industrial challenges. We invite researchers and practitioners to contribute new scenarios, particularly in:

• Asset Classes: Turbines, HVAC systems, Pumps, Transformers, CNC Machines, Robotics, Engines
• Task Domains: Prognostics and Health Management, Remaining Useful Life (RUL) estimation, Root Cause Analysis (RCA), Diagnostic Analysis, Predictive Maintenance

How to contribute:

1. Define your scenario following our Utterance Guideline and Ground Truth Guideline
2. Explore the Hugging Face dataset for examples
   • For external public sources and starter asset-class mappings, see External Industrial Dataset Guide
3. Submit a Pull Request or open an Issue with the tag new-scenario
4. Contact us with questions:
   • Dhaval Patel — [email protected]
   • Nianjun Zhou — [email protected]

---

Infrastructure Support

Model API access for AssetOpsBench 2.0 is enabled by TokenRouter (PaleBlueDot AI), a unified API platform providing access to leading AI models through a single API endpoint.

---

Contributors

Thanks to these wonderful people ✨

---

Star History

---

---