PAC-Bayes and robotic 3D printing: how to make controllers that learn from data more reliable
A new statistical method developed at Eindhoven University of Technology could improve the reliability of machine learning-based control systems in additive machines. The research addresses the problem of generalization: ensuring that a controller optimized with limited data continues to work correctly in new conditions.
The problem of generalization in control systems
Modern additive machines increasingly integrate sensors and adaptive algorithms. But how can we ensure that a controller trained on limited data also works in unforeseen conditions?
Domagoj Herceg from the Department of Mechanical Engineering at Eindhoven University of Technology published in June 2026 a study titled “PAC-Bayesian Certificates for Quadratic Closed-Loop Control”. The work proposes a method to associate statistical guarantees with the performance of controllers learned from data.
The research does not directly concern a new 3D printer. It belongs to the general theory of control. However, its implications for the additive sector are significant.
- PAC-Bayesian method to certify the performance of data-based controllers
- Applicable to additive systems with sensors and adaptive controls
- Addresses the problem of generalization beyond training data
- Published by Eindhoven University of Technology in June 2026
Why it is relevant for additive manufacturing
Modern additive machines incorporate sensors, cameras, vision systems, and thermal controls. These systems must react autonomously to varying conditions during the build.
Current 3D printers integrate algorithms capable of modifying parameters during the build. They monitor vibrations, temperature, material variations, and deposition errors. When an algorithm must react autonomously, it is not enough to know that it worked in previous tests.
One must understand how likely it is to continue behaving correctly under conditions not present in the training data set. This is precisely the problem addressed by the PAC-Bayesian approach.
The limitation of controllers optimized on data
A controller can achieve excellent results during training, but this does not guarantee equivalent performance when the system encounters new disturbances.
An algorithm can be optimized using dozens or hundreds of experimental trials. During training it can achieve very good results. This, however, does not guarantee that the same performance is maintained under different conditions.
In the case of an industrial robot, tool mass, arm position, joint friction, or vibrations can change. In an additive manufacturing machine, the variables are even more numerous.
Temperature, environmental conditions, material viscosity, extruder flow rate, component geometry, amount of accumulated heat, and mechanical behavior of the structure can vary during construction.
The PAC-Bayesian approach
The PAC-Bayes method combines Bayesian statistics and PAC (Probably Approximately Correct) theory to certify controller performance with statistical guarantees.
The approach studied by Herceg allows determining whether a controller optimized with limited data will continue to work correctly when encountering different conditions. The method provides statistical certificates on expected performance.
This type of guarantee is particularly relevant for critical applications. In large-scale robotic 3D printing, for example, a control error can compromise large components with significant costs.
Implications for the future of additive manufacturing
As additive machines become more autonomous, the need to certify the reliability of adaptive control systems will become increasingly pressing.
Herceg's work belongs to general control theory, it was not developed specifically for additive manufacturing. However, the proposed principles could be applied to control systems of additive machines that use machine learning.
The research highlights a crucial theme: the integration of learning algorithms into production systems requires not only high performance, but also guarantees on their reliability in real operating conditions.
article written with the help of artificial intelligence systems
Q&A
Who developed the PAC-Bayes method for 3D controllers?
The method was developed by Domagoj Hercec from the Department of Mechanical Engineering at Eindhoven University of Technology. The study was published in June 2026.
What is the main problem addressed by TU Eindhoven's research?
The research addresses the generalization problem in machine learning-based control systems. The goal is to ensure a controller functions correctly even under new conditions not present in the training data.
How does the PAC-Bayesian method apply to additive manufacturing?
The method provides statistical guarantees on the performance of controllers managing sensors and adaptive controls in additive machines. It is useful when algorithms must autonomously react to vibrations, temperatures, or variable deposition errors.
Why is optimization on training data insufficient?
A controller may achieve excellent results during training but fail when encountering new perturbations such as mass changes or different environmental conditions. The PAC-Bayes method certifies the probability of success beyond the initial dataset.
Does the proposed method directly concern a new 3D printer?
No, the research does not directly concern a new 3D printer but belongs to general control theory. However, its implications are significant for improving the reliability of existing additive systems.
