Efficient Data Collection Strategy for Modeling the Dynamics of Floating Robotic Vehicles Using Neural Networks

As the hydrodynamic model of a surface vehicle is highly nonlinear and of high dimension, an extensive set of experiments is required to estimate the parameters of the model. Typically, the dynamic behavior of a surface vehicle is investigated through experimental tests such as planar motion mechanisms or free running tests in specialized test facilities. However, the cost of such tests is very high, and it is challenging to design effective and efficient test conditions, especially when the model is represented by numerous parameters such as neural networks. In this paper, we propose an efficient data collection strategy for modeling the dynamics of a floating robotic vehicle using neural networks. For generalized modeling, it is important to collect diverse data by exploring all the reachable state space of the vehicle. To achieve this, a data collection policy is built based on sampling-based planning toward reducing the uncertainty of the neural network-based model. The proposed method allows for collecting more comprehensive data than other commonly used random-based data collection policies. We show experimentally that the proposed method enables more accurate modeling of both fully actuated and under-actuated floating robotic vehicles, as demonstrated by the effective execution of various tasks.