Physics-informed Neural Networks (PINNs) for Orientation Estimation from IMU sensors

Physics-informed Neural Networks (PINNs) are a special kind of universal function approximator in the form of a Deep Neural Network with its solution space constrained via the underlying physical laws of the system and are particularly suited for solving inverse problems. In this work, we develop a proof-of-concept algorithm that leverages the power of PINNs to solve the inverse problem of orientation estimation from sparse and noisy IMU sensor data (accelerometer and gyroscope). The PINN framework involves fusing a surrogate Neural Network with the underlying sensor model represented as quaternion-based differential equations. Unlike conventional methods, PINNs do not require sensor noise and bias parameters, nor do they involve numerical differentiation or integration of noisy data. Derivatives are computed elegantly via automatic differentiation within the framework. PINNs have found wide success in domiains ranging from fluid mechanics to plasma physics and epidemiology to computational imaging for astrophysics. Motivated by their broad applicability, we aimed to extend the application of PINNs to motion tracking in this study. Our preliminary investigations show promising agreement between PINNs and the ground truth orientation for data segments dominated by low frequency content. Fine-tuning the PINN architecture to handle high-frequency features, consideration of neural representations natural for quaternions, investigating of longer temporal sequences, comparison with respect to various traditional methods and efforts to reduce computational cost are under further investigation.