You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

3new

300 native recorded trajectories for sweep_objects, hook_pull, and cone_ramp. No pour scene, obsolete videos, contact labels, model weights or environments are included.

Scene Episodes Observations Train / validation
sweep_objects 100 93,935 90 / 10
hook_pull 100 82,161 90 / 10
cone_ramp 100 63,969 90 / 10

Each scene has success / partial / failure = 60 / 20 / 20. Train: 54 / 18 / 18. Validation: 6 / 2 / 2. Splits are whole trajectories.

Download and render

Rendering code and installation commands: otherWM/3new. Download 3new-data-v1.tar.gz (142,218,109 bytes). Unpack under otherWM/3new/; the archive creates data/. The payload is approximately 214 MB uncompressed.

cd otherWM/3new
bash setup.sh
.venv/bin/hf auth login
.venv/bin/python src/download_data.py
bash check.sh
bash render.sh --scene all --devices 0,1

This dataset remains private; use an account with repository access. No permission changes are made by these commands.

Contents

data/scenes/<scene>/<train|val>/<episode>/trajectory.hdf5 preserves original HDF5 bytes. All original numerical states, actions, joints, camera matrices, timestamps and native metadata are retained. Units: metres / radians / seconds; quaternion order wxyz. There are T actions and T+1 observations. Original embedded metadata can contain provenance paths, which are never followed as runtime file references.

geometry.json is the sole geometry binding for rendering: body-local meshes, recorded world poses and native physical parameters. Correct native cylinder/capsule axes are represented here; the numeric recording is NOT modified to fit an old video. native_assets.json retains original asset provenance; do not substitute those old exported visuals for the common geometry interface. data/assets/ stores shared geometry and Panda material GLBs by content hash. data/files.json contains per-file sizes and SHA256 checksums.

Code renders saved states only: no physics replay, recollection, interpolation or skipped frames. Output style is fixed at 960x720, Blender 4.5.3 / Cycles64, original frame rate, H264 CRF18. Raw input remains immutable. Rendered MP4s and validation receipts are written separately under output/.

Attribution

Panda assets originate from ManiSkill. The upstream ManiSkill license accompanies this dataset as ManiSkill-LICENSE. Third-party components retain their own rights; no blanket dataset license is asserted here. The bundle has been checked for all 300 native records, geometry links, cameras and time alignment. These checks are not a claim that all 300 new videos have already been rendered.

Downloads last month
-