Using Physics-informed Learning for Nonlinear System Identification of Underwater Robots
Accurate dynamics models are crucial for simulation, control design and state estimation for robotic systems. In the case of underwater robots and AUVs in particular, the dynamics can be highly nonlinear — this makes it difficult to simulate, control and estimate their motion in complex tasks such as docking, inspection and obstacle avoidance. Data-driven approaches can help identify nonlinear dynamics models for AUVs. Such identified models can be useful in applications including adaptive Model Predictive Control and Extended Kalman Filters. In this tutorial, we will present a workflow using JAX and physics-informed learning to learn a dynamics model from AUV pose data (positions, orientations, linear and angular velocities). We will teach the audience how to perform simulations to validate this model against a ground truth. We will also share the code for this tutorial in an open repository.