Visit the DexJoCo2 section of the DexJoCo project website for demonstrations of all 10 tasks, showing task execution alongside ten-finger tactile maps.

Demo videos

Task execution (upper panel) and ten-finger tactile maps (lower panel).

USB Insertion	Card Swiping	Power Screwdriving	Pipette Transfer	Childproof Bottle
Cabinet Retrieval	Safe Unlocking	Hammer Nail	Fold Glasses	Pick Bucket

View all demos on the DexJoCo website →

Dataset overview

Version 1.0.0 contains 2,200 trajectories across 10 tasks, with 1,154,612 frames recorded at 30 FPS:

Source	Per task	Total
Human-teleoperated demonstrations	20	200
Automatically generated extensions of the human demonstrations	200	2,000
Combined	220	2,200

The data was collected in Isaac Sim using Panda arms and Sharpa hands. RGB images, tactile indentation maps, and contact wrenches are simulated. The dataset includes multi-view RGB, robot joint states and wrist poses, tactile signals, task instructions, and timestamps.

This release replaces the earlier 200-trajectory preview and already includes those 200 human demonstrations. Do not combine it with a previously downloaded copy of the preview, as that would duplicate the human data.

Tasks

Each task is distributed as one archive containing 220 episodes.

Task	Archive
Cabinet Retrieval	`cabinet_retrieval.zarr.zip`
Card Swiping	`card_swiping.zarr.zip`
Childproof Bottle	`childproof_bottle.zarr.zip`
Fold Glasses	`fold_glasses.zarr.zip`
Hammer Nail	`hammer_nail.zarr.zip`
USB Insertion	`insert_usb.zarr.zip`
Pick Bucket	`pick_bucket.zarr.zip`
Pipette Transfer	`pipette_transfer.zarr.zip`
Power Screwdriving	`power_screwdriving.zarr.zip`
Safe Unlocking	`safe_unlocking.zarr.zip`

Format and loading

The archives contain FTP-1-compatible Zarr v2 datasets for use with Open-X-Tactile. See the FTP-1 data-processing documentation for field definitions, coordinate conventions, and loading guidance.

DexJoCo2/
├── cabinet_retrieval.zarr.zip
├── card_swiping.zarr.zip
├── ...
├── safe_unlocking.zarr.zip
└── SHA256SUMS.txt

Download the archives from the data directory and extract them before loading. Each archive contains a <task>.zarr/ directory. Episode boundaries are stored in meta/episode_ends.

RGB: ego, left-wrist, and right-wrist views; JPEG XL compression with distance=1, decoding to uint8 at the stored resolution. The reader must support and register the JPEG XL codec.
Other numeric arrays: lossless storage.
Tactile maps: SharpaWaveTacMapSim, with tactile type matrix.
Contact wrenches: SharpaWavePhysXContactSim, with tactile type state.

SHA256SUMS.txt provides checksums for the 10 archives. On Linux, after downloading all archives and the checksum file into the same directory, run:

sha256sum -c SHA256SUMS.txt

Episode origin and lineage

Within each task, using zero-based episode indices:

0–19: human demonstrations, with synthetic=false.
20–219: synthetic extensions, with synthetic=true.

Per-episode attributes preserve episode_id for the current recording, source_episode_id for its direct source, and root_episode_id for the original human seed. To find the corresponding human demonstration, match root_episode_id within the same task to an episode with synthetic=false. Multiple synthetic episodes can share a human seed while retaining distinct episode IDs.

Project and contact

Website and demos: dexjoco.github.io
OXT data format and dataset index: open-x-tactile.github.io
Contact: wangxiangyu2025@ia.ac.cn

Downloads last month: 27

Total file size:

151 GB