Abstract In this article, it is studied how well inertial sensor-based human activity recognition models work when training and testing data sets are collected in different environments. Comparison is done using publicly open human activity data sets. This article has four objectives. Firstly, survey about publicly available data sets is presented. Secondly, one previously not shared human activity data set used in our earlier work is opened for public use. Thirdly, the genaralizability of the recognition models trained using publicly open data sets are experimented by testing them with data from another publicly open data set to get knowledge to how models work when they are used in different environment, with different study subjects and hardware. Finally, the challenges encountered using publicly open data sets are discussed. The results show that data gathering protocol can have a statistically significant effect to the recognition rates. In addition, it was noted that often publicly open human activity data sets are not as easy to apply as they should be.