End-To-End Low-Level Neural Control of an Industrial-Grade 6D Magnetic Levitation System
Philipp Hartmann, Jannick Stranghöner, Klaus Neumann
AI summary
Problem
Controlling 6D magnetic levitation systems is inherently challenging due to complex, unstable dynamics and model mismatch, traditionally requiring extensive, expert-driven modular control engineering that is brittle and time-consuming.
Approach
The authors train an end-to-end recurrent neural network via behavior cloning on interaction data from a proprietary controller, directly mapping raw Hall sensor readings and 6D reference poses to coil current commands while respecting strict real-time inference constraints.
Key results
- First end-to-end neural controller for industrial 6D magnetic levitation
- Achieves robust, accurate control with bounded-error stability under reference pose jumps
- Strong generalization and extrapolation to unseen poses and system dynamics
- Real-time inference library optimized for 250 µs cycle times on industrial hardware
Why it matters
Demonstrates the practical feasibility of learning-based neural control for complex, unstable industrial systems, offering a potential paradigm shift from manual control engineering to data-driven automation.
Abstract
Magnetic levitation is poised to revolutionize in- dustrial automation by integrating flexible in-machine product transport and seamless manipulation. It is expected to become the standard drive technology for automated manufacturing. However, controlling such systems is inherently challenging due to their complex, unstable dynamics. Traditional control approaches, which rely on hand-crafted control engineering, typically yield robust but conservative solutions, with their performance closely tied to the expertise of the engineering team. In contrast, learning-based neural control presents a promising alternative. This paper presents the first neural controller for 6D magnetic levitation. Trained end-to-end on interaction data from a proprietary controller, it directly maps raw sensor data and 6D reference poses to coil current commands. The neural controller can effectively generalize to previously unseen situations while maintaining accurate and robust control. These results underscore the practical feasibility of learning-based neural control in complex physical systems and suggest a future where such a paradigm could enhance or even substitute traditional engineering approaches in de- manding real-world applications. The trained neural controller, source code, and demonstration videos are publicly available at https://sites.google.com/view/neural-maglev.