H-JEPA datasets

Data for the H-JEPA code release (hierarchical JEPA world models). Image columns are stored with lossless Blosc-Zstd compression inside the HDF5 files; h5py with hdf5plugin reads them directly. The repository mirrors $STABLEWM_HOME, the data root the code reads from, so a download with --local-dir $STABLEWM_HOME puts every file where the code expects it.

FourRoom Distractors

The four-room navigation environment with one distractor agent that wanders and teleports (Poisson, mean interval 35 steps), rendered at 224x224. The controlled agent follows a noisy expert policy in half of the episodes and a random exploration policy in the other half. Every episode has 401 frames.

hf download jepa-world-models/h-jepa --repo-type dataset --local-dir $STABLEWM_HOME --include "fourroom_tp35_d1*"

Path	Episodes	Size	Used for
fourroom_tp35_d1.h5	3,750	4.1 GB	training (the configs read 1.0M of its 1.5M transitions)
fourroom_tp35_d1_val.h5	150	0.2 GB	validation; source of the planning-evaluation tasks
fourroom_tp35_d1_probing.h5	600	0.7 GB	probe training
fourroom_tp35_d1_probing_val.h5	150	0.2 GB	probe evaluation

All four files come from one collection seed, so the validation and probing files are the first 150 / 600 / 150 episodes of the training file. scripts/data/collect_datasets.sh in the code release regenerates them (identical in our checks).

Visual AntMaze

Rollouts of the OGBench AntMaze expert policies in the medium maze, rendered at 64x64 from the back camera. Every episode has 402 frames.

hf download jepa-world-models/h-jepa --repo-type dataset --local-dir $STABLEWM_HOME --include "visual_antmaze_medium_*"

Path	Episodes	Size	Used for
visual_antmaze_medium_explore_stitch_train.h5	25,000 (12,500 explore + 12,500 stitch)	22.0 GB	training
visual_antmaze_medium_stitch_val.h5	1,250 stitch	1.4 GB	validation; source of the planning-evaluation tasks
visual_antmaze_medium_probing_train.h5	3,106 (1,555 explore + 1,551 stitch)	3.4 GB	probe training
visual_antmaze_medium_probing_eval.h5	310 (155 explore + 155 stitch)	0.3 GB	probe evaluation

The training file is the training data of the seed-42 paper models and holds only the columns training reads (pixels, action, xy, qpos, qvel, proprio, plus the episode index columns); the other files keep every column the collector records except observation, a copy of pixels. scripts/data/collect_datasets.sh in the code release collects files from the same distribution (not byte-identical).

Push-T and OGBench Cube

The training data of these two environments is LeWM's release: quentinll/lewm-pusht and quentinll/lewm-cube (MIT). This repository adds only pusht_expert_val.h5 (21 episodes, 16 MB, with the block_ori column), the Push-T validation set, converted from DINO-WM's Push-T validation data. The code release's README lists the download and preparation commands (add_pusht_block_ori.py, split_cube.py).

DROID

A copy of DROID 1.0.1 prepared for H-JEPA training and the offline DROID planning evaluation.

hf download jepa-world-models/h-jepa --repo-type dataset --local-dir $STABLEWM_HOME --include "droid/*"
bash $STABLEWM_HOME/droid/extract.sh            # add --delete to remove each tar once extracted

extract.sh checks SHA256SUMS, then untars every shard in place (JOBS=8 in parallel). The loader (h_jepa/droid_data.py) resolves relative paths (the CSVs, their entries and the manifest's episode_path) against $STABLEWM_HOME/droid.

Path	Content
droid/droid_paths_minus16_256p.csv	training split: 74,896 episodes (<episode dir> <index>, relative paths)
droid/droid_val_indist_256p.csv	validation split used for monitoring: 64 episodes (62 of them also in the training CSV)
droid/droid_256p/shard-{00000..00060}.tar	61 tar shards (90.9 GB, 0.56-1.52 GB each): the 256x256 training corpus, extracting to droid/droid_256p/1.0.1/<lab>/<success|failure>/<date>/<episode>/
droid/droid_raw_eval16.tar	0.25 GB: the 16 raw DROID 1.0.1 episodes of the planning-evaluation manifest, extracting to droid/droid_raw/1.0.1/...
droid/SHA256SUMS, droid/extract.sh	tar checksums and the extraction script

Training corpus (droid_256p). DROID 1.0.1 re-encoded to 256x256 mp4, all episodes minus the 16 evaluation episodes. Each episode directory holds the files the loader reads: the DROID metadata json, trajectory.npz (the cartesian_position, gripper_position and camera_extrinsics/* arrays of DROID's trajectory.h5) and the two exterior-camera mp4s (recordings/MP4/<serial>.mp4, the left_mp4_path and right_mp4_path views of the metadata; the wrist camera is not included). The mp4 containers are tagged 60 fps but hold DROID's 15 Hz footage; the models use every third frame (5 fps). Totals: 74,516 episode directories, 295,718 files, 90.4 GB. As in the paper runs, 382 CSV entries have no episode directory and 1,191 episodes lack one or both exterior mp4s; the loader replaces a clip that fails to load by another random episode, so these entries are kept to reproduce the paper's sampling exactly.

Evaluation episodes (droid_raw). The 16 episodes read by the planning-evaluation manifest (h_jepa/droid_assets/droid_clips_waypoint_curated16v2_5fps_gw36.json), as in the raw DROID 1.0.1 release (1280x720 mp4): per episode the metadata json, trajectory.h5 and the left exterior-camera mp4 (48 files).

License and attribution. DROID is released by its authors under CC BY 4.0; this re-encoded subset is redistributed under the same license. The changes are the 256x256 re-encoding, the trajectory.npz extraction and the file selection above. Please cite DROID when using this data:

@inproceedings{khazatsky2024droid,
  title     = {{DROID}: A Large-Scale In-The-Wild Robot Manipulation Dataset},
  author    = {Khazatsky, Alexander and Pertsch, Karl and Nair, Suraj and Balakrishna, Ashwin and
               Dasari, Sudeep and Karamcheti, Siddharth and Nasiriany, Soroush and Srirama, Mohan Kumar
               and Chen, Lawrence Yunliang and Ellis, Kirsty and others},
  booktitle = {Robotics: Science and Systems},
  year      = {2024}
}

Checksums

The root SHA256SUMS lists the FourRoom, AntMaze and Push-T validation files (DROID has its own droid/SHA256SUMS): cd $STABLEWM_HOME && sha256sum -c --ignore-missing SHA256SUMS.