DexJoCo-v1.1-preprocessed

PointWAM's training frames for DexJoCo v1.1 (Franka + Allegro, MuJoCo, 11 tasks), re-rendered from the official v1.1 demos: RGB-D at 224 x 224 with body ids, camera poses, robot state and actions, one H5 per demo. The same data in LeRobot v3 is chrockey/DexJoCo-v1.1-preprocessed-lerobot. MIT (License).

Download and check

hf download chrockey/DexJoCo-v1.1-preprocessed --repo-type dataset --local-dir DexJoCo-v1.1-preprocessed
cd DexJoCo-v1.1-preprocessed && sha256sum -c --quiet MANIFEST.sha256

Pin a commit of this repo with --revision <commit> for reproducible runs. Add --include "water_plant/*" for one task (then check with grep ' water_plant/' MANIFEST.sha256 | sha256sum -c --quiet).

Source

• Demos: DexJoCo/DexJoCo-Datasets-v1.1, revision 077bbed4a3c9807ae88dad06842b7989533ab2bc, group rand_obj, all 1,100 episodes (11 tasks x 100).
• Simulator: brave-eai/dexjoco tag v1.1 = a02e9cf (task XMLs and env classes used as released).
• Code: PointWAM cdcfdd6b (pointwam/preprocessing/dexjoco/README_v11.md): replay_capture_v11 (pass 1, replay with upstream's reset and stop rule) and extract_frames.py --dexjoco-version 1.1 --size 224 (pass 2, GPU EGL render), then scripts/trim_dexjoco_start_artifacts.py.

Settings

setting	value
image size	224 x 224, all cameras
anti-aliasing	MSAA off (offsamples=0, as v1.1's XMLs and videos)
screen effects	on: ipad_screen_effect (unlock animation) and camera_screen_effect (photograph live viewfinder, freeze + flash after the shutter), re-rendered per frame
stop rule	first success (upstream's); hammer_nail keeps 30 frames after success
trimmed demos	5 demos whose recording is frozen at the start: click_mouse episodes 17 and 78, bimanual_microwave_cook 72, 30, 31; the frames before the replay rejoins the recording are dropped (first kept frame 143, 171, 376, 177, 98; attrs trim_start, trim_reason)

Every demo succeeds on replay.

Contents

task	arms	demos	frames	cameras	GB
bimanual_assembly	2	100	57,876	back, handcam_rgb_left/right	12.4
bimanual_hanoi	2	100	109,046	back, handcam_rgb_left/right	24.9
bimanual_microwave_cook	2	100	69,788	back, handcam_rgb_left/right	16.2
bimanual_photograph	2	100	40,293	back, handcam_rgb_left/right	7.9
bimanual_unlock_ipad	2	100	47,094	back, handcam_rgb_left/right	10.2
click_mouse	1	100	39,658	front, handcam_rgb	5.2
fold_glasses	1	100	57,694	front, handcam_rgb	8.3
hammer_nail	1	100	25,034	front, handcam_rgb	4.2
pick_bucket	1	100	46,841	front, handcam_rgb	6.9
pinch_tongs	1	100	43,381	front, handcam_rgb	5.6
water_plant	1	100	27,470	front, handcam_rgb	4.1
total		1,100	564,175		106.0

GB = 10^9 bytes. Frame counts are after trimming.

Layout <task>/<source episode name>.h5, named after the v1.1 episode's source recording (bimanual_unlock_ipad: v1.1's own re-recordings, bimanual_unlock_ipad_demo_<k>_replay_...). stats/ holds PointWAM's normalization stats for this 11-task corpus at 224: dexjoco-v11-all11-224/norm_stats.json and the siblings dexjoco-v11-all11-224_{robotflow,action,action_2arm}.json (_action single-arm, _action_2arm bimanual); pass stats/dexjoco-v11-all11-224 as data.norm_stats_path.

Differences from the v1.0-era chrockey/DexJoCo-preprocessed: demos from v1.1 (corrected bimanual_unlock_ipad initial states and actions, 4 changed water_plant demos), episodes end on the replay's success frame (v1.1's rule; several tasks are shorter than in v1.0, e.g. bimanual_assembly 57,876 vs 72,642 frames), MSAA off instead of 8x, screen effects rendered, and water_plant's spray cone rendered.

H5 schema (T frames at 30 Hz)

key	shape	meaning
attrs		task, prompt, n_arm, n_frames, size, success, cameras, dexjoco_version, source_group, source_episode_index, source_episode_path; trim_start, trim_reason in trimmed demos
actions	(T, 23·n_arm)	per arm [pos3, quat4, hand16], arms interleaved
proprio	(T, 23·n_arm)	[tcp pose 7·n_arm | hand joints 16·n_arm] (blocked)
qpos	(T, nq)	MuJoCo qpos
body_names, is_robot, is_hand	(B,)	bodies and robot / hand masks
xpos, xquat	(T, B, 3/4)	body poses
<cam>/rgb, /depth, /body_id	(T, 224, 224, ·)	RGB, depth (m, float16), body id
<cam>/intrinsic, /extrinsic	(3, 3), (T, 4, 4)	camera intrinsics, per-frame pose

source_episode_index is the episode index in v1.1's rand_obj/<task> dataset; source_episode_path is that episode's orig_episode_path from its replay_initial_values.jsonl, relative to upstream's raw-dataset root (<task>/<recording name>).

Use with PointWAM

Copy stats/* into the PointWAM checkout's stats/, then train on the 11 task folders, e.g. the multi-task fine-tune:

python train.py experiment=finetune_dexjoco_mt_volt_h data=dexjoco/mt11_all1100 \
    'data.data_dirs=[<root>/bimanual_assembly,<root>/bimanual_hanoi,...,<root>/water_plant]' \
    data.norm_stats_path=stats/dexjoco-v11-all11-224

Manifest

MANIFEST.sha256 (sha256sum -c format, paths relative to the dataset root) covers all 1,106 files of the dataset: every H5, stats/, this README and LICENSE. Working files of the build (pass-1 captures, partial pass-2 outputs, the 5 untrimmed originals, logs) are not part of the dataset.

License

MIT (LICENSE). Made from DexJoCo/DexJoCo-Datasets-v1.1 (MIT, Copyright (c) 2026 Dexjoco contributors) with the DexJoCo simulator (MIT). The rendered robots are the robot models bundled with DexJoCo: Franka Emika Panda (Apache-2.0) and Wonik Allegro hand (BSD-2-Clause, Copyright (c) 2016, SimLab); no model files are included here. If you use this data, please cite DexJoCo:

@misc{wang2026dexjocobenchmarktoolkittaskoriented,
      title={DexJoCo: A Benchmark and Toolkit for Task-Oriented Dexterous Manipulation on MuJoCo},
      author={Hanwen Wang and Weizhi Zhao and Xiangyu Wang and Siyuan Huang and He Lin and Boyuan Zheng and Rongtao Xu and Gang Wang and Yao Mu and He Wang and Lue Fan and Hongsheng Li and Zhaoxiang Zhang and Tieniu Tan},
      year={2026},
      eprint={2605.16257},
      archivePrefix={arXiv},
      primaryClass={cs.RO}
}