Real-Time Path Planning With Virtual Magnetic Fields

Humans and animals have learned or evolved to use magnetic fields for navigation. Knowing how to model and estimate these fields can be used for motion planning. However, computing the propagation of electromagnetic fields in a given environment requires solving complex differential equations with advanced numerical methods, and therefore it is not suitable for real-time decision making. In this latter, we present a real-time approximator for Maxwell's equations based on deep neural networks that predicts the distribution of a virtual magnetic field. We show how our approximator can be used to perform autonomous 2D navigation tasks, outperforming state-of-the-art navigation algorithms, ensuring completeness, and providing a near-optimal path up to 200 times per second without any post processing stage. We demonstrate the effectiveness of our method with physics-based simulations of an unmanned aerial vehicle, an autonomous car, as well as real-world experiments using a small off-road autonomous racing vehicle. Furthermore, we show how the approach can be applied to multi-robot systems, video game technology, and can be extended to 3D problems.