DROID-Sim release tools

Shujie Zhang, Jingkun Yi, Weipeng Zhong, Zirui Zhou, Yangkun Zhu, Hanqing Wang, Xudong Xu^†, Weinan Zhang, Chunhua Shen

^* Equal contribution. ^† Corresponding author.

Structure

release/
  package/                 maintained tools, skill, reports, final media
  batch_1st_scene/<scene>/  accepted reconstruction and original frame 0
    trajectory/            original DROID episode plus decoded replay inputs
    render/scene.json      portable rendering camera/coordinate contract
    parameters.yaml        final geometry and physics provenance
  replay/                  experiments, render checks, conversion/replay outputs

Scene content	Meaning / authority
image.png, source_image.json	Original chosen input frame and its provenance.
fully_refined_foreground.glb	Accepted foreground geometry and final placement, authority for rendering and node IDs.
sam3d/	Per-object masks/cutouts, generated assets and original composition records.
depth/, reconstruction/	Estimated depth, point clouds and original reconstruction cameras; not a metric-scale certificate.
refinement/	Retained projection/sim-ready records.
background/	Unchanged native Marble Gaussian PLY and panorama, completion image and aligned GS export. The completion image is not the rendering background.
render/scene.json	Current shared-camera coordinate contract for accepted foreground and aligned Gaussian background.
parameters.yaml, physics/colliders/	Final-node physics inputs, recovered collider registrations.
trajectory/	Byte-identical DROID originals, decoded measured states and camera-to-base calibration.
layered_composite.png	Historical image used for verification. Current real-GS previews are under release/replay.

Environment and rendering

The tested rendering Python is /cpfs/user/zhangshujie/miniconda3/envs/ikea-asset/bin/python. requirements-render.txt records the working non-Isaac dependency versions; PyTorch was 2.5.1+cu121.

From this directory, set SCENE to a full scene path, then:

python tools/render_scene.py --scene "$SCENE" \
  --output-dir "../replay/render_validation/$(basename "$SCENE" | cut -c1-4)" \
  --device cuda:0

This renders the actual Marble Gaussian PLY. The same camera observes both layers through an explicit coordinate transform.

--orbit-deg 5 rotates the same camera for both layers about the foreground's world-space bounds center and gravity Z axis. Zero degrees preserves the exact stored camera. --width 640 is optional for previews.

For a continuous orbit with assets kept in memory:

python tools/render_orbit.py --scene "$SCENE" --output-dir ../replay/orbit_preview --device cuda:0
python tools/orbit_batch.py --output-dir ../replay/orbit_renders \
  --montage media/frame0_orbits.gif --devices cuda:2 cuda:3

Defaults are +/-10 degrees, 81 frames at 10 fps (approximately 8 seconds). The batch is checks um-resumable and refuses to publish an incomplete all-scene montage. Novel views can expose missing geometry; these previews do not certify complete reconstruction. The fixed initial camera-gauge optimization was tested against independent per-frame normalization at 5 degrees (maximum 1/255 difference).

May need a little time to load the GIF.

Original trajectories

Each trajectory directory contains:

• trajectory.h5: unchanged original DROID trajectory.
• metadata_<original-uuid>.json: unchanged original DROID metadata and filename.
• replay.npz: observed and commanded joints/gripper, observed Cartesian pose, timestamps, and selected camera pose arrays. Observed joints are the default state-playback data; commands are not interchangeable with observations.
• initial_state.json: the first measured robot/gripper state.
• camera.json: selected camera identity and frame-0 camera-to-robot-base matrix.
• manifest.json: checksums, array meanings, provenance, and conventions.

USD conversion

/isaac-sim/python.sh tools/glb2usd.py --folders "$SCENE" \
  --pattern fully_refined_foreground.glb --output-dir ../replay/usd_preview

Omit --output-dir to write adjacent USD files if needed. --overwrite is explicit.

Robot Replay

Dataset limitation. Gripper and object physics parameters have not yet been tuned across all dataset scenes to ensure successful replay of DROID tasks.

Physics inputs

parameters.yaml retains collision references and records whether their source mesh can be registered to the final geometry.

Friction values are low-confidence language-model material priors. Gravity is a standard Earth-gravity default in the scene's gravity frame. Metric scale, mass, collider loading, cavities and final contacts remain explicit validation fields.

Citation

If you find our scene dataset useful, please consider citing our work using the following BibTeX entry. Thank you for your support!

@misc{zhang2026robosnaponeshotrealtosimscene,
  title={RoboSnap: One-Shot Real-to-Sim Scene Generation for Generalizable Robot Learning and Evaluation},
  author={Shujie Zhang and Jingkun Yi and Weipeng Zhong and Zirui Zhou and Yangkun Zhu and Hanqing Wang and Xudong Xu and Weinan Zhang and Chunhua Shen},
  year={2026},
  eprint={2607.06699},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2607.06699}
}