Vladyslav Fediukov, Felix Dietrich, Fabian Buse, and Jana Huhne
Paper presented at ISTVS 2024 | 21st International and 12th Asia-Pacific Regional Conference of the ISTVS Keywords: Terramechanics; Rover Locomotion; Uncertainty Quantification; Multi-Fidelity; Machine Learning https://doi.org/10.56884/6VTE9FAQ
In predicting locomotion on soft-soil, we have to deal with a consistent uncertainty surrounding this process, from the input noise to the uncertainties produced by an approximation. Available data comprises limited sets of experimental data and various numerical approximations. Machine learning models, gaining recognition in the terramechanics community, need to work with these limited data set. By deploying probabilistic frameworks, like Gaussian processes, for our tasks, we can implicitly work with the resulting uncertainties. Accurate uncertainty quantification and further analysis can provide more robustness and understanding of terramechanical machine learning models. In our work, we concentrate on the uncertainty's propagation, uncertainty calibration, and uncertainty decoupling for a wheel locomotion prediction. Our machine learning models work in a multi-fidelity framework using experimental data from the DLR’s TROLL testbed and numerical simulations using TerRA and SCM, approximating the high-fidelity target function in the training process. The experimental setup involves runs with various velocities and movement scenarios, including tilting, steering, for- and backward, as well as up- and downhill movements. Complete analysis of uncertainties will give engineers and operators a more in-depth understanding of the reliability of ongoing simulations and predictions. Moreover, uncertainties quantification can help us improve our data generation and modeling process, allowing us to make a self-improving model.
Full paper purchase: https://www.istvs.org/proceedings-orders/paper ISTVS members receive three complimentary papers per year: https://www.istvs.org/members
Last updated
