Researchers use agentic AI to monitor and correct 3D prints — system catches errors in real time, uses modular design to work on different makes and models

Researchers from the Carnegie Mellon University's Department of Mechanical Engineering just devised a system that uses multiple large language models to monitor and correct 3D printers in real time, as spotted by TechXplore. While additive manufacturing techniques have revolutionized the field, especially for customized products and prototype models, most 3D printers are still prone to errors. For example, Prusa3D said about 7% of its prints on the MMU2S failed, while a further 19% required user attention (although they did not fail).

This meant that users need to keep an eye on the printing process to ensure that everything goes as planned. While this might not be an issue for home and casual users running one or two 3D printers, it starts to be a problem if you’re using them for manufacturing. Although there’s no global standard, many manufacturers aimed for failure rates of about 5% in the 1980s. But today, the standard is closer to 0.1%, meaning a 7% failure rate is incredibly wasteful and makes 3D printing less competitive on quality than other manufacturing processes.

To fix this, the Carnegie Mellon research team used four specialized large-language model agents with a supervising agent to ensure optimization. The first agent is a visual-language model that takes photos after each printed layer and then analyzes them for print quality and defects. Another agent then looks at the current printer settings to identify what needs to be changed or improved to solve the detected issues.

The information is handed off to a solution planner agent which creates an actionable plan. This is then given to the executor agent, which interacts with the 3D printer through the API to get the desired outcome. All four agents are managed by the supervisor agent, ensuring that all the information is relevant and up to date.

What’s crucial with this system is that it does not use custom LLMs trained on a specialized dataset to work effectively. Instead, it only used the base ChatGPT-4o and domain-specific, generalized structured prompts that the team developed. This makes it simple to implement and improve 3D printing performance and efficiency.

"The future is adaptive," Associate Professor of Mechanical Engineering Amir Barati Farimani said. "The integration of LLMs into the 3D printing process represents a significant advancement. As these models evolve, their ability to reason over richer, multimodal data will unlock even more capabilities. For now, this work provides a foundation for truly intelligent and autonomous manufacturing systems, capable of achieving unprecedented levels of precision and reliability."

If this tech becomes widely adopted, perhaps the cameras on our 3D printers will eventually feed an LLM control loop instead of being used for manual babysitting. Until that future arrives, though, print failures and plastic spaghetti will have to be prevented through manual monitoring and intervention.

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.