Reducering av sim-to-real gapet för undervattensfordon med residual dynamisk modellering

Accurate simulation of underwater vehicles is essential for effective design, testing and deployment of autonomous underwater systems. However, a persistent challenge in underwater robotics is the sim-to-real gap, the discrepancy between simulation and reality. Within this domain the gap is especially pronounced in dynamics modelling. Contributing factors to this discrepancy between simulated and real-world dynamics include complex hydrodynamic effects, environmental disturbances and modelling limitations. This thesis investigates a real-to-sim-to-real approach to improve simulation fidelity by learning residual dynamics; data-driven corrections that augment an existing nominal model, in a computationally efficient manner to assure real-time simulation capability. Specifically, the study compares two machine learning regression-based methods, K-Nearest Neighbours (KNN) and Gaussian Processes (GP), for learning residual dynamics of a BlueROV2. These models are trained on real-world data collected via an underwater motion capture system in a controlled tank environment and are evaluated in terms of simulated acceleration and velocity prediction accuracy. To assure the models real-time simulation feasibility, their inference times were also evaluated. The results show that while both models reduce acceleration error, KNN consistently outperforms GP across most scenarios in terms of overall simulation accuracy. Notably, large baseline errors underscore the critical need for residual correction. However, increased velocity error in GP models introduces ambiguity in the overall benefit, highlighting the complexity of evaluating model effectiveness. The results also show that KNN is acceptable for real-time simulations but for GP it depends on its configuration on the tested computer. The study also evaluates the effect of input space complexity. While models trained on a smaller control input space had more accurate zero-shot baselines, those trained with full 6-DoF inputs achieved greater improvements, especially in the rotational axes. This suggests that model expressiveness can be better utilized in scenarios where the data sufficiently captures the system dynamics.