h2r-robots-1h30

A 1.53-hour subset of h2r-robots (6 datasets, about 15 minutes each, random episodes with seed 0), sized like the human-video subset used for the same world model so that only the video source differs. Real robot manipulation video, resampled to 4 fps and center-cropped to 224 × 224 JPEG frames. One third-person camera per dataset. Used to train a JEPA world model in human-to-robot.

In h2r-robots, episodes were sampled at random (seed 0) from the LeRobot versions of the datasets below. A video id <dataset>/episode_<index>.mp4 points to the episode index in the source dataset.

Source	Robot	Camera	License (as listed on the source)
lerobot/droid_1.0.1	Franka	exterior_1_left	Apache 2.0
lerobot/berkeley_autolab_ur5	UR5	image	CC BY 4.0
lerobot/jaco_play	Kinova Jaco	image	CC BY 4.0
lerobot/taco_play	Franka	rgb_static	CC BY 4.0
lerobot/stanford_kuka_multimodal_dataset	KUKA iiwa	image	MIT
lerobot/roboturk	Sawyer	front_rgb	Apache 2.0

Please cite the original datasets and Open X-Embodiment when using this data.

Layout

• frames/<split>-<shard>.parquet: one row per episode, video (id) and frames (list of JPEG bytes).
• manifest.parquet: video id, split, number of frames, shard.
• config.yaml: the extraction config.