DynamicTHOR: A Scalable Dataset of Human-Centric Dynamic Scenes for Embodied AI

DynamicTHOR is a synthetic dataset of human-centric dynamic indoor scenes designed for embodied AI research. The dataset simulates daily human activities and the resulting object relocations over time, enabling the study of dynamic environments beyond static scenes.The dataset consists of two main components: (1) a character activity dataset including 50 distinct personas with five-day activity schedules and associated activity–object probability models, and (2) 100 instantiated dynamic scenes with time-indexed object configurations that can be directly loaded and replayed in the ProcTHOR simulator. The data are intended to support research on embodied navigation, memory, human-aware planning, and other tasks that require reasoning about dynamic object states.Because the initial scene layouts (house_info.json in each scene folder) are derived from the ProcTHOR dataset, this record includes a copy of the Apache 2.0 license (license_procthor in DynamicSceneDataset.zip) for those files.In addition to the dataset, this record provides two accompanying codebases. The DynamicSceneGeneration module contains the reference implementation of the dynamic scene generation pipeline described in the manuscript, enabling users to generate additional personas and dynamic scenes following the same workflow. The generation code is provided for transparency and reproducibility; the released dataset corresponds to a fixed version generated using the configuration described in the manuscript. The DynamicTHOR module provides utilities and example scripts for loading and interacting with the released dynamic scenes within the ProcTHOR-RL framework.