🚀 AstroBench

Free-Floating Capture and Space Robotic Manipulation

📄 Paper · 💻 GitHub · 🌐 Project · 📖 Task catalog

AstroBench studies how robots capture drifting and rotating objects and manipulate space hardware. In free-floating capture, a contact can move the target away or change its rotation, making approach timing and contact geometry central to the task.

This Hugging Face release brings together an 8,414-object expansion inventory, five cabin scene configurations, and two editable space workcells with nine manipulation tasks and 29 task–target combinations. Object and task catalogs can be explored in the dataset viewer or loaded with 🤗 Datasets. The MELFI and RRM geometry, tools, mounting hardware, collision meshes and editable sources are available in the asset package.

MELFI sample storage	RRM servicing panel

📦 Load the catalogs

pip install datasets huggingface_hub

from datasets import load_dataset

repo_id = "brickasdasda/AstroBench"

objects = load_dataset(repo_id, "objects_objaverse8414", split="catalog")
tasks = load_dataset(repo_id, "tasks", split="catalog")
prompts = load_dataset(repo_id, "prompts_objaverse8414", split="catalog")

print(objects[0]["name"], objects[0]["source_url"])
print(prompts[0]["text"])

The default configuration is objects_objaverse8414. Each configuration uses a catalog split containing published definitions and records.

Configuration	Rows	Contents
objects_objaverse8414	8,414	Selected Objaverse object names, category labels, source IDs, attribution and license metadata
objects_core20	20	Core object classes and canonical prompt names
tasks	9	Workcell task definitions, target lists, initial states, measurements and scoring parameters
prompts_core20	109	Four capture tracks × 20 core objects, plus 29 workcell instructions
prompts_objaverse8414	33,685	Four capture tracks × 8,414 expansion objects, plus 29 workcell instructions
scenes	5	Cabin identifiers, placement configurations and source references
workcells	2	MELFI/RRM parts, frames, layout, asset identities and archive members

The expansion inventory preserves each object's source link and original attribution. Its 8,414 rows identify individual objects; category labels are recorded separately. Use the source IDs and links to retrieve the corresponding Objaverse models. Scene records similarly link their cabin assets to the original source.

🛠️ Space hardware and tasks

MELFI V6 — sample storage. Two narrow boxes and one wide box form a tray mounted along the cabin wall. The model includes separate lids, latch parts, retention bands and sample cards. Each of the three boxes supports five tasks:

Task ID	Instruction template
card_transfer_to_mouth	Move the card to the opening of the {target_name} and hold it aligned.
card_insert	Insert the card fully into the {target_name}, release it, and leave it retained inside.
card_extract	Remove the card from the {target_name} and hold it clear of the opening.
lid_open	Open the lid of the {target_name} and leave it open.
lid_close	Close and latch the lid of the {target_name}, then withdraw the gripper.

RRM TB1 — servicing caps and captive screws. The panel is mounted vertically on a cabin wall, with the arm's floor mounting frame directly in front of it. The geometry includes the FDV1 cap, a 3 × 3 SMA cap array, two captive fasteners, cap adapters and tool docks.

Task ID	Instruction template
fdv1_remove_stow	Unscrew and remove the {target_name}, retain it in its adapter, dock the adapter in its stowage holder, and withdraw the driver.
sma_remove_stow	Unscrew and remove the {target_name}, retain it in its adapter, dock the adapter in its stowage holder, and withdraw the driver.
captive_loosen	Loosen the {target_name} without releasing it from the panel and hold it in the loosened position.
captive_tighten	Tighten the {target_name} until it is seated, keep it captive, and withdraw the driver.

The prompt tables resolve {target_name} into each exact box, cap or screw name. Capture instructions use the same wording across the static, drift, tumble and fast_tumble tracks. The tables preserve the published registry version and text hashes.

Download and open the workcells

Download the complete workcell package

from pathlib import Path
from zipfile import ZipFile
from huggingface_hub import hf_hub_download

archive = hf_hub_download(
    repo_id="brickasdasda/AstroBench",
    repo_type="dataset",
    filename="assets/astrobench-workcells.zip",
)
with ZipFile(archive) as z:
    z.extractall(Path("astrobench-assets"))

The archive contains the full workcells/ directory, including both workcells and their shared source files. Open workcells/melfi_v6/melfi_v6.blend or workcells/rrm_tb1/rrm_tb1.blend in Blender, or import the corresponding .glb in a compatible renderer. Parts, collision data, dimensions and transforms are described in the adjacent JSON files. The package uses metres and radians.

The accompanying workcell guide describes task resets, target selection, world placement and the arm mounting frames. Use the GitHub package for episode planning, task scoring and result aggregation.

🎬 Mechanism demonstrations

FDV1 removal and stowage · SMA cap removal and stowage · Captive screw

These archived demonstrations show the mechanisms in their original tabletop arrangement. The overview images and packaged layout describe the current wall-mounted arrangement. Demonstration metadata records the individual videos and layout review.

📐 Data fields and provenance

• Object records preserve stable object_id values; Objaverse records also retain source_uid, source_url, author attribution and the original license label.
• Prompt records contain task_id, target_id, track, text, and hashes for the instruction and source registry.
• Task and workcell records expose identifiers and commonly used scalar fields directly. Heterogeneous specifications are stored as canonical JSON in fields ending in _json; read them with json.loads().
• Workcell archive member paths locate the GLB, Blender, parts, collision and manifest files. The original manifests retain source and asset checksums.

This release is derived from AstroUpLab/AstroBench at commit 66a8e1f. See provenance.json for the source revision and file checksums.

📜 License and sources

AstroBench's original task definitions, code and reconstructed workcell geometry are released under Apache-2.0. Third-party object assets retain the individual licenses recorded in the object catalog. Their author attribution and source links are preserved. NOTICE and the workcell package's own NOTICE describe the NASA/ESA hardware references and asset provenance.

📚 Citation

If you use AstroBench in your research, please cite:

@misc{wang2026astrobench,
  title         = {{AstroBench}: Benchmarking Free-Floating Capture and Space Robotic Manipulation},
  author        = {Zizhe Wang and Jia Deng and Xin Lu and Zehao Jia and Yi Huang and Yao Zhu},
  year          = {2026},
  eprint        = {2610.31064},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  doi           = {10.48550/arXiv.2610.31064},
  url           = {https://arxiv.org/abs/2610.31064}
}