Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('validation'): ('json', {}), NamedSplit('test'): (None, {})}
Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

OrbitBench

100 numbered spacecraft assets, four presentation videos, and a six-page research manuscript. This release packages the exact frozen Blender revisions used for the videos, together with corresponding simulation previews.

These are staged candidate models, not certified high-fidelity or training-ready reconstructions. See verification.json and the limitations below.

Selected surface model, asset 1

Videos

Video Duration Corresponding Models
Earth-orbit fleet 18 s All 100
White-studio reveal 24 s All 100
Grid scan and component cascade 24 s 32-model subset
Satellite collision presentation 11 s 16-model subset

All videos are 1920 x 1080, 24 fps, silent, and have no overlaid names. video/manifest.json gives the exact numbered membership of each video. Presentation scales, orbits, and collision animations are illustrative, not ephemeris propagation or validated impact physics.

Asset Layout

assets/1/ ... 100/
  1.blend           Original video-source Blender model; all textures packed
  1.glb             Original assembled base-geometry GLB
  1.xml             Numbered MuJoCo preview entry
  1.usda            Numbered USD physics asset entry
  scene.usda        Standalone zero-gravity USD scene
  models/           Supporting USD layers and preview definitions
  components/       Component meshes required by the simulation
  preview.png       Selected original surface render
  metadata.json     Release identity, original identity, and source hashes
  provenance/       Original surface and geometry records

Replace 1 with the desired asset number. Numbering follows the frozen video manifest: 1 is Hubble, 2 is the ISS, and 3 is TESS. This is a naming change, not anonymization. Original identities and component-level names remain in the mapping table and provenance.

Open assets/<n>/<n>.blend in Blender 4.5.3 or a compatible newer version for the exact surface model used in the videos. The file is byte-identical to the frozen source and includes all packed textures. Original scene metadata and render-output paths are retained; set a local output directory before writing new renders.

Use assets/<n>/<n>.xml for MuJoCo, or assets/<n>/scene.usda for a standalone USD scene. Reference <n>.usda when composing a larger simulation. USD prim identifiers use Satellite_<n> with numeric display names. Keep each numbered directory intact: its simulation dependencies resolve inside that directory.

The GLB, USD, and MuJoCo files preserve the base geometry and preview materials. Detailed procedural surface shaders are in Blender; the other formats are not claimed to reproduce those shaders exactly.

Existing preview exports were copied for 97 assets. Missing preview definitions for Landsat 9 (81), QZS-1R (83), and EarthCARE (90) were generated only in this release using the existing project exporter and original component meshes. Their metadata identifies this supplementation. No source geometry was rebuilt.

Paper And Provenance

paper/main.pdf is the unchanged six-page IEEE RA-L-style research draft. Its statistics describe the September 24, 2026 snapshot, not newly certified results. Its spacecraft names correspond to mapping.csv. The individual quality reports and large combined fidelity PDFs are not the research manuscript and are not duplicated in this package.

provenance/ retains the original video manifest, source receipts, and status documents. Historical references to local-only delivery describe the original snapshot. This repository publishes the packaged snapshot without changing its research acceptance decisions.

Verification And Limitations

verification.json records 100 MuJoCo load/30-step free-body checks, 100 USD composition/dependency checks, self-contained GLB checks, full decoding of the four videos, and a six-page manuscript check.

These checks are not a visual inspection, Isaac Sim runtime test, engineering accuracy evaluation, contact validation, or training-readiness certification. Mass, inertia, and collision proxies remain unverified. The original final high-fidelity acceptance count is zero.

The repository owner's original MIT license metadata is retained. Third-party imagery, credits, and source-specific restrictions remain applicable; this release does not claim to relicense third-party source material.

After downloading the repository contents, verify the delivered files from the repository root:

sha256sum --check --quiet SHA256SUMS

Local .cache/ upload-resume files are not part of the release or checksum inventory and are not uploaded.

Downloads last month
104