SmartPilot: A Multiagent CoPilot for Adaptive and Intelligent Manufacturing
SmartPilot: A Multiagent CoPilot for Adaptive and Intelligent Manufacturing
Motivation Behind Smart Copilot
In modern manufacturing, AI solutions tend to focus on isolated problems. We might rather have one system that detects anomalies in machinery, another that forecasts production delays, and perhaps a third that reasons over logs or documents. But these are typically disjoint systems, they don’t talk to each other or reason over shared context.
This fragmented approach causes major blind spots. For example, a defect detection model may trigger alerts, but if it lacks causal reasoning, operators won’t understand why the issue happened or how to prevent it in the future. Likewise, a predictive system may forecast delays but not explain them.
SmartPilot is an agent-based copilot that addresses these limitations by reasoning holistically over time-series data, causal dependencies, and operator knowledge. It’s built to support real-time decision-making while being explainable, compact, and extensible..
Overview of the SmartPilot framework.
At the heart of the system is a multiagent architecture, consisting of four intelligent agents:
PredictX, which monitors data streams and predicts anomalies before they occur;
ForeSight, which models time-based dependencies and forecasts future production outcomes;
CausalLens, which maps causal chains and discovers root causes of failures; and
InfoGuide, a chatbot-style assistant that answers domain-specific questions grounded in manuals or logs.
These agents are not isolated, they interact, share intermediate outputs, and even cross-reference predictions. For example, PredictX can trigger CausalLens when an anomaly is detected, and InfoGuide can provide corrective instructions grounded in historical context.
The system also features multimodal data integration — sensor data, images, and documentation are all fused in a standardized internal schema.
Finally, the neurosymbolic core enables agents to combine statistical learning with symbolic rules. This is key for explainability, which is critical in industrial AI applications.
Uniqueness Of the Project
It introduces SmartPilot, a modular multiagent system for intelligent manufacturing. Each agent specializes in a domain, such as anomaly prediction, causal reasoning, or document Q&A but also collaborates via a shared architecture.
Second, we developed innovative algorithms for each agent. These are lightweight models designed for real-time factory environments, integrating both neural and symbolic elements.
Third, we implemented a multiagent architecture that allows decentralized task execution while maintaining inter-agent memory synchronization. This mimics a collaborative team of AI experts supporting human operators.
Fourth, we incorporated neurosymbolic AI methods to improve explainability without sacrificing performance. This hybridization lets us reason over events like “if machine A has a lag and temperature is high, then failure probability increases” , something pure deep learning often struggles with.
Finally, we ran comprehensive evaluations, including both quantitative benchmarks and qualitative expert reviews, to ensure the system is practical and reliable in real-world manufacturing contexts.
Link to the Paper, github-repo, slides, demo-video
This work was awarded the best paper award for its application in manufacturing