External Wrench Estimation for Multirotor Aerial Vehicles

This work presents the development and analysis of the Knowledge-Based Neural Ordinary Differential Equation (KNODE) methodology to model multirotor aerial vehicles (MRAV), integrating the physics-based model derived from Newton-Euler formalism with a neural network to capture unmodeled residual dynamics. The primary objective is to enhance the accuracy of the vehicle’s external wrench estimation problem, especially in scenarios where the first-principle model dynamics are uncertain. This framework addresses one of the limitations of the state-of-the-art model-based wrench estimators: the wrench estimation of these observers comprises the external wrench (e.g. collision, physical interaction, wind) in addition to residual wrench (e.g. model parameters uncertainty or unmodeled dynamics). This is a problem if these wrench estimations are to be used as wrench feedback to a force controller, for example. To train the KNODE model, it is assumed that both the twist and the wrench of the base are known and available to be analyzed offline; these quantities can be estimated in practice fusing Internal Measurement Unit (IMU) data with optitrack technology, and if available, collecting force-torque sensor data. To integrate the learned KNODE model into the external wrench observer framework, the momentum-based observer is implemented to estimate the external wrench acting on the MRAV comprehending of the learned residual dynamics. This hybrid approach balances first-principle accuracy with the flexibility of data-driven learning. The method is validated through numerical simulations of an aerial robot in various flight scenarios and residual dynamics. Statistical analysis of the results indicates that wrench estimation can improve by up to 70% with minimal effort, leaving room for further enhancement.