README.md

metadata

pretty_name: DexJoCo Single-Arm Image Data
license: mit
language:
  - en
tags:
  - robotics
  - dexjoco
  - behavior-cloning
  - image-observations
  - single-arm
source_datasets:
  - DexJoCo/DexJoCo-Datasets-Raw

DexJoCo single-arm image datasets

Processed demonstrations for all six single-arm DexJoCo tasks, ready for the DICE-RL image behavior-cloning loader. This public release contains 596 episodes and 216,890 retained transitions. Four demonstrations were removed before processing due to poor trajectory quality: pick_bucket demo 39 (244967) and demo 12 (974934), click_mouse demo 42 (581035), and hammer_nail demo 2 (403013). They were not replaced. This 2026-10-21 release contains 98 pick_bucket, 99 click_mouse, 99 hammer_nail, and 100 episodes for each other task. Every remaining raw recording passed conversion; no additional episodes were rejected. Per-task normalization for the three affected tasks was recomputed using only the clean retained episodes. No policy checkpoints or measured policy success rates are included.

Task	Episodes	Transitions	Retained episode length	Archive GiB
click_mouse	99	31,928	192–538	1.01
fold_glasses	100	53,878	282–844	1.91
hammer_nail	99	21,041	149–392	0.99
pick_bucket	98	41,829	233–725	1.73
pinch_tongs	100	40,469	288–680	1.10
water_plant	100	27,745	203–394	1.10

Total archive size: 7.84 GiB. Allow approximately 15.7 GiB for cached archives plus extracted data.

Download

Use the updated scripts in stpark7/dice-rl, with huggingface_hub installed in the activated Python environment:

bash script/download_hf.sh

No token or Hugging Face login is required. The destination is $DICE_RL_DATA_DIR/dexjoco if that variable is set, otherwise the checkout's data_dir/dexjoco. To select tasks or a destination:

bash script/download_hf.sh --task click_mouse --task water_plant --data-dir /path/to/data

The optional dexjoco selector is supported. The script resolves one commit for the whole download, verifies each archive's SHA-256, and installs each task atomically. Record the printed commit or use --revision <commit> to reproduce a release. Existing unrelated datasets are preserved; use a new data directory if the target task directory already exists without a matching download receipt. A repeat download checks the archive and receipt, but does not re-audit edited local arrays. Original Robomimic downloads use bash script/download_hf.sh robomimic.

data_dir/dexjoco/<task>-img/
├── episodes.json
├── ph_pretrain/
│   ├── train.npz
│   └── normalization.npz
└── ph_finetune/
    ├── train.npz
    └── normalization.npz

Each NPZ contains states, actions, traj_lengths, camera_names, and images_<camera> for every recorded camera. Finetuning additionally includes rewards and terminals. Images are embedded in both training files, so each NPZ is self-contained. This is a custom NPZ/tar layout rather than a datasets.load_dataset table. The complete file index, sizes, checksums and per-task statistics are in manifest.json.

Data and conversion

• All cameras are retained in this order: ego_left, ego_right, front, wrist. Their arrays are RGB uint8, shaped (T, 3, 96, 96).
• States are float32, 23-dimensional [tcp_pose7, hand_joints16]. Pose quaternions use wxyz ordering.
• Actions are float32, 22-dimensional: world translation delta (3), world relative rotation vector (3), and absolute hand-joint targets (16). The rotation increment left-multiplies the measured TCP orientation.
• States and actions use per-task min/max normalization to approximately [-1, 1] with epsilon 1e-6, fitted on all retained episodes. There is no reserved validation/test split. Images are not normalized by the loader.
• Leading zero-command markers are removed. Initial held-target prefixes are trimmed after sustained target departure: position 0.001 m, rotation 0.5°, hand joints 0.01 rad, 3 consecutive samples, minimum held duration 0.5 s, retained context 0.1 s. Interior pauses remain intact.
• MP4 frame t is assumed to align with recorded state/action sample t. Every camera's decoded frame count must equal the original state count; the same prefix cut is applied to all modalities. The recordings do not independently establish camera synchronization through timestamps.
• Episode ordering, original/retained lengths, prefix cuts and thresholds are recorded in episodes.json. Published source paths are relative to the raw dataset root.
• ph_finetune uses the same sequences and images as ph_pretrain, with a synthetic reward/terminal of 1 at every retained episode end and 0 elsewhere. Recorded task-success labels are not imported. These synthetic rewards do not certify demonstration or policy success.

Choose cameras explicitly in train_dataset.image_keys, for example [front, wrist]; the loader decompresses only those camera members. Keep this order consistent with the environment, model num_img and image shape. To keep a previous local installation, download this release to a new directory:

bash script/download_hf.sh --data-dir data_dir/dexjoco-clean-20261021

The older mixed single-arm/bimanual repository is a separate release and is unchanged by this publication.

Verification

The three affected tasks were freshly converted from the cleaned local raw recordings using the same converter defaults. Unaffected tasks reuse the prior verified NPZ files; all six tasks were audited again against the available raw recordings before packaging:

• All 596 episodes' stored states and actions were compared exactly against fresh raw re-encoding, the recorded prefix cuts and the published normalizers. The encoder is shared with the converter; this is a consistency check rather than an independent action-encoding implementation.
• All shared arrays in both training files, including 867,560 retained camera frames, matched exactly. Shapes, dtypes, finite values, episode boundaries, normalization extrema and synthetic rewards were checked. No all-black camera frames were found.
• 216 camera frames were compared pixel-for-pixel with fresh raw MP4 decoding: first/middle/last retained frames in first/middle/last episodes, for every task and camera. There were no mismatches.
• Each task passed a CPU BC-loader sample check using [front, wrist], an eight-step action horizon and one conditioning frame.
• Converter, action-contract and downloader regression tests passed. Archive SHA-256 values are verified by the downloader before extraction.

Conversion/audit environment: Python 3.11.15; numpy 1.26.4, scipy 1.15.3, zarr 2.18.3, av 18.1.0, torch 2.7.0+cu128, huggingface_hub 0.36.2. Exact converter/video/action/loader source-file SHA-256 values are recorded in the manifest because the release was produced from an edited working tree.

Source and license

Original demonstrations: DexJoCo/DexJoCo-Datasets-Raw. Simulator and source documentation: brave-eai/dexjoco. Training/conversion code: stpark7/dice-rl, derived from DICE-RL.

The upstream raw dataset is published under the MIT license. The upstream copyright and license notice are included in LICENSE.

Please cite the original DexJoCo and DICE-RL work when using their data/code:

• DexJoCo: A Benchmark and Toolkit for Task-Oriented Dexterous Manipulation on MuJoCo.
• From Prior to Pro: Efficient Skill Mastery via Distribution Contractive RL Finetuning.