Abstract
The successful deployment of deep learning-based techniques for autonomoussystems is highly dependent on the data availability for the respective systemin its deployment environment. Especially for unstructured outdoorenvironments, very few datasets exist for even fewer robotic platforms andscenarios. In an earlier work, we presented the German Outdoor and OffroadDataset (GOOSE) framework along with 10000 multimodal frames from an offroadvehicle to enhance the perception capabilities in unstructured environments. Inthis work, we address the generalizability of the GOOSE framework. Toaccomplish this, we open-source the GOOSE-Ex dataset, which contains additional5000 labeled multimodal frames from various completely different environments,recorded on a robotic excavator and a quadruped platform. We perform acomprehensive analysis of the semantic segmentation performance on differentplatforms and sensor modalities in unseen environments. In addition, wedemonstrate how the combined datasets can be utilized for different downstreamapplications or competitions such as offroad navigation, object manipulation orscene completion. The dataset, its platform documentation and pre-trainedstate-of-the-art models for offroad perception will be made available onhttps://goose-dataset.de/. \