Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Electromagnetic tracking via Deep Learning
The goal of this project is to work on a new magnetic tracking system that combines our expertise in machine learning, haptics, and electromagnetic actuators. The outcome will be incorporated as one of the building blocks of a larger system we are working on that is oriented towards VR/AR haptics.
Keywords: Machine Learning, Haptics, Electromagnetic Actuators, Finite Element Method, FEM.
In this project, we seek to track in real-time the position of a permanent magnet incorporated into a pen, based on magnetic sensor readings. We recently published the first generation of the haptic system [1]. In this explorative work, we demonstrated the viability of the actuation concept. However, for tracking we relied on an external Optitrack system. We are now working on a self-contained device that incorporates a magnetic-based tracking to the proven-to-work magnetic actuation. in order to completely eliminate the reliance on an external tracking system.
Previous approaches for solving for a magnet position in the 3D space were based on a simplified version of the full electromagnetic model. The full model is unfeasible to solve in real-time. The simplified approach works well under certain assumptions: the magnets being tracked are spheres, perfectly magnetize, and the magnetic sensors are far enough apart from each other [2]. If these assumptions are not met, or ferromagnetic materials are present, the model fails. These assumptions severely limit the accuracy and restrict possible applications of our tracking system. Therefore we want to loosen these assumptions by incorporating more of the full electromagnetic model with the help of deep learning. More specifically, we seek a neural network that can approximate the Finite Element Method (FEM) model simulation of the full electromagnetic model. In general, the idea of combining well-established physical models with machine learning is gaining attention in various fields, such as in fluidics [3], optics [4], and mechanical analysis [5].
We are seeking a motivated student to develop this hybrid computation model that takes advantage of a fast analytical model, for where the postulates are met, combined with data-driven correction terms learned with the help of an accurate FEM generated dataset. This includes several magnets geometries and non-linear magnetic materials. The differences to the analytical model will be learned in a deep learning approach, and then used in the tracking process. Once the dataset is created, different architectures, losses, and approaches to learning the general function can be explored. Depending on the progress and the student interest, the next step will be to optimize sensor placement via an evolutionary strategy approach and to get involved in the hardware development of the next generation of the haptic prototype.
[1] https://ait.ethz.ch/projects/2020/magnipulator/
[2] https://doi.org/10.1109/JSEN.2019.2936766
[3] https://doi.org/10.1145/2816795.2818129
[4] https://doi.org/10.1021/acsnano.9b02371
[5] https://doi.org/10.1109/ACCESS.2020.2977880
In this project, we seek to track in real-time the position of a permanent magnet incorporated into a pen, based on magnetic sensor readings. We recently published the first generation of the haptic system [1]. In this explorative work, we demonstrated the viability of the actuation concept. However, for tracking we relied on an external Optitrack system. We are now working on a self-contained device that incorporates a magnetic-based tracking to the proven-to-work magnetic actuation. in order to completely eliminate the reliance on an external tracking system.
Previous approaches for solving for a magnet position in the 3D space were based on a simplified version of the full electromagnetic model. The full model is unfeasible to solve in real-time. The simplified approach works well under certain assumptions: the magnets being tracked are spheres, perfectly magnetize, and the magnetic sensors are far enough apart from each other [2]. If these assumptions are not met, or ferromagnetic materials are present, the model fails. These assumptions severely limit the accuracy and restrict possible applications of our tracking system. Therefore we want to loosen these assumptions by incorporating more of the full electromagnetic model with the help of deep learning. More specifically, we seek a neural network that can approximate the Finite Element Method (FEM) model simulation of the full electromagnetic model. In general, the idea of combining well-established physical models with machine learning is gaining attention in various fields, such as in fluidics [3], optics [4], and mechanical analysis [5].
We are seeking a motivated student to develop this hybrid computation model that takes advantage of a fast analytical model, for where the postulates are met, combined with data-driven correction terms learned with the help of an accurate FEM generated dataset. This includes several magnets geometries and non-linear magnetic materials. The differences to the analytical model will be learned in a deep learning approach, and then used in the tracking process. Once the dataset is created, different architectures, losses, and approaches to learning the general function can be explored. Depending on the progress and the student interest, the next step will be to optimize sensor placement via an evolutionary strategy approach and to get involved in the hardware development of the next generation of the haptic prototype.
Each year the IDEA League offers the students of its partner universities over 180 monthly grants for a short-term research exchange. In general, these grants are awarded based on academic merit. For more information visit http://idealeague.org/student-grant/
Master Thesis
CLS Student Project [managed by Max Planck ETH Center for Learning Systems]