Abstract

Agentic AI is poised to transform decision-making in Industry 4.0 by enabling autonomous agents to reason over multimodal inputs—such as sensor streams, structured knowledge bases, and unstructured maintenance logs—and act adaptively under uncertainty. Yet, real-world adoption remains challenging due to data fragmentation, integration complexity, limited explainability, and lack of evaluation workflows.

This hands-on tutorial offers a full lifecycle walkthrough for building trustworthy agentic AI systems in industrial settings. Participants will engage in two interactive labs: (i) resolving data silos in smart manufacturing using an open-source platform, and (ii) benchmarking agent performance, reasoning, and explainability in an enterprise-scale industrial simulation. Capabilities such as trace visualizations, real-time introspection, and comparative reasoning will be demonstrated. The session concludes with best practices for governance, monitoring, and reusable evaluation workflows. Participants will leave with practical skills and modular tools to build explainable, robust, and deployable agentic AI systems for real-world Industry 4.0 applications.

Introduction

Agentic AI is rapidly becoming a cornerstone of intelligent decision-making in Industry 4.0. These agents must reason across heterogeneous data sources—including sensor time series, structured knowledge graphs, and unstructured logs—while adapting under uncertainty. Despite advances in large language models and multimodal learning, building deployable and trustworthy systems remains difficult due to fragmented data, lack of explainability, and limited evaluation protocols.

This tutorial provides a comprehensive, hands-on walkthrough of the lifecycle of multimodal agentic AI—from design to deployment—featuring lab sessions on data integration and benchmarking. We explore reasoning strategies, evaluation methods, and governance tools that ensure trustworthy and auditable deployments.

Goals of the Lab

• Understand the full lifecycle of multimodal agentic AI systems in Industry 4.0.
• Learn key agent architectures and reasoning strategies, including Plan-Execute, ReAct, Reflexion, RAFA, and hybrid symbolic-neural methods.
• Gain hands-on experience integrating multimodal data (sensors, text, knowledge graphs).
• Diagnose integration challenges such as data silos and misaligned modalities.
• Explore evaluation methods using human feedback, sensor-grounded truth, and LLM-as-a-judge approaches.
• Apply governance and monitoring techniques for agent traceability, explainability, and coordination.

Schedule

Technical Setup

Participants should bring a laptop with Python 3 installed. Pre-configured environments and setup instructions will be provided. The tutorial uses:

• Python 3
• Google Colab or Jupyter Notebooks
• Hugging Face for LLM-driven agents
• LLM access for agent execution
• Simulation environments: AssetOpsBench, SmartPilot, FailureSensorIQ, FutureFactories dataset
• Optional Docker setup for isolated local execution

Supplementary Materials

1. AAAI-2025 Lab Video
2. Anomaly Detection Code Pattern
3. Trustworthy AI @ SWITCH AIMX 2024
4. Multi-Agent Systems for Enterprise: TechXchange 2025
5. AssetOpsBench
6. AssetOpsBench Challenge @ CODS-COMAD 2025
7. SmartPilot CoPilot System
8. FailureSensorIQ Dataset
9. Smart Manufacturing Wiki
10. Neurosymbolic AI Wiki

References

• Patel et al. 2025. AssetOpsBench: Benchmarking AI Agents for Task Automation. arXiv:2506.03828
• Constantinides et al. 2025. FailureSensorIQ: A Multi-Choice QA Dataset. arXiv:2506.03278
• Shyalika et al. 2025. SmartPilot: Multiagent CoPilot for Manufacturing. arXiv:2505.06492
• Harik et al. 2024. FutureFactories Dataset. arXiv:2401.15544
• Patel et al. 2024. AI Model Factory: Scaling AI for Industry 4.0. AAAI Proceedings

Organizers

Chathurangi Shyalika

Ph.D. student at AIISC, University of South Carolina. Research in Deep Learning, Multimodal-AI, Neurosymbolic-AI, anomaly detection, event understanding.

Saumya Ahuja

AI Engineer Lead, IBM WatsonX ASEAN. Leads Generative AI and Agentic AI projects across APAC. Experienced in LLMs, RAG systems, and enterprise AI deployments.

Shuxin Lin

Researcher at IBM with expertise in AI for Industry 4.0, agent evaluation, multimodal reasoning, and large-scale industrial AI benchmarks.

Ruwan Wickramarachchi

Research Scientist at Bosch Center for AI. Ph.D. from AIISC, USC. Research in Generative AI, Neurosymbolic AI, knowledge graphs, and multimodal representation learning.

Dhaval Patel

Senior Technical Staff Member, IBM Research. Expert in Data Mining, Machine Learning, Time Series, and industrial AI platforms such as Maximo and AutoAI-TS.

Amit Sheth

NCR Chair & Professor, AIISC, USC. Fellow of IEEE, AAAI, ACM, AAAS. Research in trustworthy, explainable, and safe neuro-symbolic AI.