BEHAVIOR-100 includes a dataset of 500 VR demonstrations. Each BEHAVIOR-100 activity is demonstrated five times: three times in the same scene and instantiation, showing variation from execution; once in the same scene with a different instantiation; and once in a different scene. Demonstrators in VR have embodiment identical to one of BEHAVIOR-100’s embodiments

Virtual reality demonstration dataset (VR_demos)

This processed dataset could be useful for imitation learning dataset, as described in the BEHAVIOR documentation. It includes all observation modalities (semantic segmentation, instance segmentation, RGB, depth, highlight), task observations, proprioception, and the action used by the agent. Additionally, each hdf5 contains metadata on the specific task.

To download a single example demonstration (~1 gb):

wget https://download.cs.stanford.edu/downloads/behavior/bottling_fruit_0_Wainscott_0_int_0_2021-05-24_19-46-46_episode.hdf5

To download the entire dataset (~250 gb)

wget https://download.cs.stanford.edu/downloads/behavior/behavior_imitation_learning_v0.5.0.tar.gz

Raw virtual reality demonstration dataset (VR_demos_raw)

To download the raw hdf5s, which include eyetracking and individual pose of the VR sensors, please use the following command:

To download a single example demonstration (~10 mb):

https://download.cs.stanford.edu/downloads/behavior/bottling_fruit_0_Wainscott_0_int_0_2021-05-24_19-46-46.hdf5

To download the entire dataset (~1.7 gb)

wget https://download.cs.stanford.edu/downloads/behavior/behavior_virtual_reality_v0.5.0.tar.gz