The Robot@VirtualHome dataset is a raw collection of data from 30 virtual homes with different appearance obtained through the Robot@VirtualHome ecosystem. Each virtual house imitates a real house, keeping the same room layout and imitating real objects with virtual object models. The objectives of this dataset are: first, to be used as a testbed for diverse algorithms such as semantic mapping through the categorization of objects and/or rooms, active exploration of the environment, localization by appearance, or others where the data presented are of interest, and second to provide a basic example of the results that can be obtained through the Robot@VirtualHome ecosystem. The dataset consists of 113278 captures in 30 houses, with 236 rooms, 2569 objects and 4 different appearance conditions. Each data capture has stored an RGB image, a depth image, a semantic mask image, the measurements from a laser scanner and a log with information about the position at which the data was taken. In addition, for each house we have added the occupancy map obtained with the laser scanner and a log with the ground truth of all objects and rooms. Five raids have been carried out for each house: the first one, capturing data at the nodes of a grid using standard appearance, in the remaining four raids the data were taken by wandering around visiting all the rooms and using different appearance conditions. More detailed information is provided in the article. An API is available here to facilitate access to the dataset data.

This work has been supported by the research projects WISER (DPI2017-84827-R), funded by the Spanish Government and financed by the European Regional Development's funds (FEDER), arpeggio (PID2020-117057GB-I00), funded by the European H2020 program, and the UG PHD scholarship program from the University of Groningen.