End-to-End Low-Level Neural Control of an Industrial-Grade 6D Magnetic Levitation System

Abstract

Magnetic levitation is poised to revolutionize industrial 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 demanding real-world applications. The trained neural controller, source code, and demonstration videos are publicly available at https://sites.google.com/view/neural-maglev.

Figures (7)

• Figure 1: Traditional engineering approaches (red box) involve expert teams that hand-craft control pipelines typically consisting of numerous modules based on a system model. However, 6D magnetic levitation is complex and therefore, model assumptions may differ from the real system -- known as model mismatch. Our neural control approach (green box) can inherently capture all systematics by learning from interaction data from the system in an end-to-end fashion.
• Figure 2: The defined workspace (yellow) illustrates all valid mover poses over a one-tile system in $x$ and $y$. Data is generated from responses of a proprietary controller to lift-off trajectories $\tilde{\mathbf{p}}'_{n}(t)$, each with constant, randomly sampled values for $x, y, \alpha, \beta, \gamma$. Lift-off and stable hovering mainly affects the $z$-axis, shown on the right. To improve robustness, all dimensions are augmented by reference pose jumps, consisting of a perturbation phase (red, duration $\Delta t_{n,\mathrm{pert}}$) followed by a correction phase (blue, duration $\Delta t_{n,\mathrm{corr}}$).
• Figure 3: Average translational and rotational responses to 100 randomized reference pose jumps, validating bounded-error stability for the proprietary controller $\pi_{PC}$ (red), the neural controller $\pi_{NC}$ (blue), the same neural controller with residual correction network $\pi_{NC} + \mathcal{C}$ (green). Standard deviations are shown as shaded areas.
• Figure 4: Accuracy measurements for a circular trajectory tracking experiment regarding the proprietary controller $\pi_{PC}$ (red), the neural controller $\pi_{NC}$ (blue), the same neural controller with residual correction network $\pi_{NC} + \mathcal{C}$ (green), and the reference circle with a radius of $100µm$ (dashed black).
• Figure 5: The neural controller $\pi_{\text{NC}}$ remains stable while hovering despite carrying payloads of varying masses and mass distributions.