HDVIO2.0 was developed to overcome the limitations of conventional visual-inertial measurement (VIO), which degrades performance in the presence of low-availability vehicle models and persistent external disturbances such as wind. It introduces a six-degrees-of-freedom (6DoF) translational and rotational vehicle dynamics model, tightly integrating it with VIO while minimizing computational load in real-time applications. It captures complex aerodynamic effects using a hybrid dynamics model combining a point-mass vehicle model and learning-based components, and accesses control commands and IMU history to represent rotational dynamics as a continuous time function. It utilizes the difference between actual and predicted motion to estimate external forces and robot state. It outperforms state-of-the-art methods on publicly available and novel drone dynamics datasets and in real-world flight experiments with winds up to 25 km/h. It demonstrates that accurate vehicle dynamics predictions are possible even without precise knowledge of the vehicle state.
