x-square-robot/x2real-assets

X2Real Simulation Assets

This repository distributes a versioned snapshot of simulation assets used by X2Real ID evaluation and local simulation workloads. It contains robot and object models, indoor scenes, materials, textures, lighting resources, grasp labels, and task asset catalogs and pools.

This repository contains static simulation assets for X2Real, not LeRobot episodes, observations, actions, videos, or trajectories. Simulator code, task definitions, control logic, and policy weights are provided separately.

Restored asset tree

The payload inventory contains 74,423 files:

Directory	Files	Contents
navigation/	70,359	3DSceneData scenes, USD scenes and objects, textures, maps, and lighting; about 34.44 GB
benchmark/	3,455	Evaluation objects, scene backgrounds, and HDRI lighting; about 8.81 GB
art_assets/	329	Models, scenes, materials, URDF, and STL resources
grasp_labels/	196	Object grasp-label arrays (.npy)
robots/	45	Franka and ARX R5 USD models and related materials
simready/	4	EX001 robot USD versions and sensor USD configuration
manipulation/	12	Task object categories and attribute JSON files
objects/	11	Brick, paper, and other USD objects and textures
pools/	11	Task sampling pools for bottles, dishes, fruit baskets, drawers, and mobile manipulation
render_settings/	1	green_booth render settings

The main formats are USD/USDA, PNG/JPG, EXR, MDL, JSON, NPY, and a small number of URDF, STL, GLB, PGM, and YAML files.

Evaluation coverage

The snapshot covers 5 boards and 73 unique ID cases:

Board	Setting	Cases
artixon_arm6a_tabletop	ArtiXon Arm-6A tabletop manipulation	38
artixon_arm6a_challenge	ArtiXon Arm-6A challenge tasks	10
quanta_x1_mobile	Quanta X1 mobile manipulation	3
arx_r5_tabletop	ARX R5 tabletop manipulation	12
franka_panda_tabletop	Franka Panda-Hand tabletop manipulation	10

Repository layout

README.md                         # This dataset card
SHA256SUMS                        # SHA-256 checksums for distribution files
upload-manifest.json              # Paths, sizes, and SHA-256 for all 107 TAR shards
restore_assets.py                 # Restore and verify the original asset tree
scripts/
  download_x2real_assets.py        # Download, verify, and optionally restore
bundle/
  README.md                       # Original snapshot notes
  manifest.json                   # Source, version, counts, tree digest, and checks
  inventory.jsonl                 # Per-file path, size, and SHA-256
shards/
  assets-001-of-107.tar
  ...
  assets-107-of-107.tar

Each shard is an independent TAR archive. All shards together restore the same asset root. Do not concatenate the shards with cat. The repository can show its README before every shard is available, so check upload-manifest.json and the checksums before restoring.

Download and restore

Download this public Hugging Face dataset:

python3 -m pip install -U huggingface_hub
hf download x-square-robot/x2real-assets scripts/download_x2real_assets.py --repo-type dataset --local-dir ./x2real-assets-tools
python3 ./x2real-assets-tools/scripts/download_x2real_assets.py --output ./x2real-assets

The helper downloads all 107 shards into ./x2real-assets-download, verifies every file listed in SHA256SUMS, and restores the asset tree into ./x2real-assets. It requires Python 3.12 or newer. To keep the download and restore steps separate, use:

python3 ./scripts/download_x2real_assets.py --download-dir ./x2real-assets-download
python3 ./x2real-assets-download/restore_assets.py --output ./x2real-assets

The output directory must be absent or empty. Point the asset-root option of your simulator to ./x2real-assets; do not point it at the download directory or at shards/. The restore script checks TAR sizes and SHA-256 values, then verifies the payload file count, directory count, file sizes, per-file SHA-256 values, and the complete tree digest.

Keeping both the downloaded shards and the restored tree requires about 88.3 GB before filesystem overhead; reserve at least 100 GB of free space. After restoration, the same checks can be run again:

python3 ./x2real-assets-download/restore_assets.py --output ./x2real-assets --verify-only

Dataset viewer

For metadata browsing in the Hugging Face Dataset Viewer, use asset_catalog.csv. It contains one row per payload file with its relative path, type, byte size, and SHA-256. The simulator assets remain packaged in the shards/ directory and are restored with the download script.

Project links

• Paper: X2Real: An eXtensive Simulation Benchmark for Real-World Generalist Policies
• Code: X-Square-Robot/x2real
• License: TODO