VA-Bench

VA-Bench evaluates embodied spatial intelligence through visual demonstrations, active perception, and metric robot control. It contains 14 physics-simulated manipulation task families: 11 single-arm tasks and three dual-arm tasks.

• Paper: VA-Bench
• Code and benchmark: github.com/zhangzhongbo2213/VABench
• Upstream simulator: RoboTwin
• Upstream assets: RoboTwin2.0

Contents

This repository distributes the versioned simulator assets needed by the 14 main VA-Bench tasks. The archive contains the ALOHA-AgileX robot description, planning and collision configurations, and every object variant selected by the vabench_eval and demo_clean task configurations.

File	Purpose
va-bench-main-assets-v1.tar.gz	Complete main-task asset archive
va-bench-main-assets-v1.tar.gz.sha256	Archive checksum
ASSET_MANIFEST.json	Per-file size and SHA-256 manifest
VALIDATION.json	Release validation record
data/tasks.jsonl	Structured index of the 14 task families
data/assets.jsonl	Structured index of the 189 bundled asset files

The archive expands to an assets/ directory. It does not contain the full RoboTwin asset collection; randomized backgrounds and clutter require the upstream assets.

Download the simulator assets

The binary archive is a simulator resource bundle rather than a table of training examples. Download it with huggingface_hub:

from huggingface_hub import hf_hub_download

archive = hf_hub_download(
    repo_id="zhangzhongbo2213/VA-Bench",
    repo_type="dataset",
    filename="va-bench-main-assets-v1.tar.gz",
    revision="assets-v1",
)
print(archive)

Extract it from the root of a VA-Bench checkout:

tar --keep-old-files -xzf /path/to/va-bench-main-assets-v1.tar.gz -C /path/to/VABench
python /path/to/VABench/script/check_setup.py

The expected archive SHA-256 is:

1a264fd776997b3dc32e4ba5d2334a00d117790b80afa851ac2bdbc31ce5f4bf

Browse the metadata

The Dataset Viewer and datasets library expose metadata indexes; they do not materialize the simulator asset tree:

from datasets import load_dataset

tasks = load_dataset(
    "zhangzhongbo2213/VA-Bench", "tasks", split="train"
)
assets = load_dataset(
    "zhangzhongbo2213/VA-Bench", "assets", split="train"
)

Each task row records the active arm, action horizon, demonstration seed, and 20 physically verified evaluation seeds. Each asset row records its archive path, byte size, and SHA-256 digest.

Validation

The release was validated from an independently extracted bundle. All 14 task environments reset successfully, rendered 1280x960 RGB observations, executed a camera action, and wrote replay videos. All 18 selected object variants loaded, rendered, and stepped. See VALIDATION.json for the complete record. These are environment checks, not model-performance evaluations.

License and provenance

The selected assets originate from RoboTwin2.0 and are distributed under the MIT License. The upstream copyright and license are retained in LICENSE, and the source and modification notice is retained in NOTICE.md.

Geometry and textures are unchanged. Only two cuRobo configuration files were modified to replace machine-specific asset paths with portable relative paths.

Citation

@article{zhang2026vabench,
  title   = {VABench: Measuring Embodied Spatial Intelligence through Visual Demonstrations, Active Perception, and Metric Control},
  author  = {Zhang, Zhongbo and Jin, Jiayi and Wang, Yifan and Zhang, Zaibin and Diao, Haiwen and Wang, Lijun and Lu, Huchuan},
  journal = {arXiv preprint arXiv:2609.19554},
  year    = {2026}
}