Official-Space-AI/physgraspgen-dataset

PhysGraspGen soft-object grasps 1.0.0rc1

Simulated grasps of soft objects, with the physics behind every frame. A Universal Robots UR7e arm with a Robotiq 2F-140 gripper picks up deformable objects in a model of a robot cell. Each frame holds the images of a fixed front camera and a wrist camera together with what the simulator computed at that step: contact forces, the stress and deformation gradient in every tetrahedron of the object's mesh, segmentation and depth.

Frames	933,896
Episodes	2,486 (2,307 successful grasps and 179 failures, both kept)
Object types	5
Cameras	front, wrist (568 x 320)
Channels	rgb, depth, force, stress, deformation_gradient, segmentation
Frame rate	30 fps, recorded at 60 fps
Format	physgraspgen-public 2.0 (specification)
Code	OpenPhysGraspGen 1.0.0rc1 (Apache-2.0)
License	CC-BY-NC-4.0
DOI	10.5281/zenodo.23787395

Load

With the Hugging Face datasets library, one config per table:

from datasets import load_dataset

frames = load_dataset("Official-Space-AI/physgraspgen-dataset", "frames", streaming=True)
trajectories = load_dataset("Official-Space-AI/physgraspgen-dataset", "trajectories", streaming=True)

The RGB images are AV1 videos under videos/, one per camera and episode. To get the files as published and read an episode with its videos:

hf download Official-Space-AI/physgraspgen-dataset --repo-type dataset --local-dir physgraspgen-dataset
pip install "openphysgraspgen[dataset]"

from openphysgraspgen.dataset import episode_metadata, read_frames, read_trajectory, read_video

root = "physgraspgen-dataset"
uid = "sponge-s26266211-d3-e000"
frames = read_frames(f"{root}/data/sponge/{uid}.parquet")
trajectory = read_trajectory(f"{root}/trajectories/sponge/{uid}.parquet")
rgb = read_video(f"{root}/videos/front/sponge/{uid}.mp4")  # (N, H, W, 3)
meta = episode_metadata(f"{root}/data/sponge/{uid}.parquet")

read_frames restores each column's shape: images are (N, H, W), per-tetrahedron values (N, T). Video frame i is the episode's frame i. Decoding needs an ffmpeg with an AV1 decoder, such as libdav1d, on the PATH.

Check

The script that tests this dataset ships in it. From the downloaded folder, with any integer as the seed:

cd physgraspgen-dataset
python -I tools/check_dataset.py . --seed 12345 --json report.json

It hashes every file and reads only those whose sha256 is the one metadata.yaml states, then tests four items: the frame count (scale), every value metadata.yaml states against the files (metadata), that every frame's RGB, depth, force and stress come from one simulation step and that contact points projected into the images land on the object or the robot (consistency, on a sample drawn from your seed), and the object types and the spread of the scene variables (diversity). A run passes only with the script and the acceptance values in tools/ that metadata.yaml lists. It prints a table, ends with a RESULT line, and exits 0 when every item passed. It needs Python 3.10 or newer with NumPy, PyArrow and PyYAML, which pip install "openphysgraspgen[dataset]" installs. See The check for what each item applies.

PG-T1

PG-T1 asks whether a grasp lifts the object, holds it and puts it down. It is scored from this dataset alone, on a CPU:

openphysgraspgen benchmark pg-t1 physgraspgen-dataset --json pg-t1.json

Scores on this release (task search_lift_hold_carry_place_10cm, all 2,486 episodes, 95% intervals):

Criterion	Lifted	Held	Placed	Task score
C, 5 cm for 2 s	2,486/2,486, 1.000 [0.998, 1.000]	2,484/2,486, 0.999 [0.997, 1.000]	2,307/2,486, 0.928 [0.917, 0.938]	0.976
B, 12 cm for 4 s	0/2,486	0/2,486	0/2,486	0.000

Under C, 29 episodes dropped the object and 150 did not place it on the mat. No grasp in this release lifts the object 12 cm, so none reaches B yet. Later releases add episodes that lift and hold the object to B's height and time.

The rules are in docs/benchmark/pg-t1.md.

Files

README.md                                   this card
.gitattributes                              keeps the hashed files byte for byte in git
metadata.yaml                               counts, object types, conventions, every file with its sha256
SHA256SUMS                                  sha256 of every data and tool file
tools/                                      the check script and its criteria
data/<object_type>/<episode_uid>.parquet    one row per published frame, every channel but RGB
trajectories/<object_type>/<episode_uid>.parquet
                                            every recorded simulation step: robot and object state, phase
videos/<camera>/<object_type>/<episode_uid>.mp4
                                            RGB, AV1

Each episode's metadata (material category, placement, lighting and camera settings, label and the PG-T1 task block) is JSON in its Parquet footer under the key physgraspgen. The format specification lists every column with its shape and unit.

Object types

Type	Episodes	Frames	Geometry
sponge	444	167,293	150 x 90 x 30 mm block, yellow
rubber_gasket	450	166,439	60 mm ring, black
silicone_pad	408	153,818	80 x 40 x 25 mm pad, blue
puff_pastry	424	159,917	80 x 50 x 30 mm block, cream
tofu	760	286,429	85 x 110 x 45 mm packaged block, cream with a pale green lid

The dataset is updated regularly: later releases add episodes and new object types. This release is checked with criteria that ask for 5 object types and 933,896 frames, so the check's report says that its criteria are weaker than the dataset test's, which asks for 10 object types and 1,000,000 frames.

How it was made

Generated with SpaceAI's simulator on the Newton physics engine (VBD solver). Contact forces are read from the solver's contact buffers, and stress comes from the deformed mesh, per tetrahedron, through the object's hyperelastic material model. Material values are taken from the literature or chosen in short simulation runs; they are not measurements of the objects shown, and only their category labels are published. Every object shape was built by SpaceAI from its dimensions, and every object colour was set by SpaceAI. metadata.yaml records the code versions that generated, exported and finalized the files.

Limitations

The data is simulated. The images are renders of a model of the cell, not camera frames, and the forces and stresses are the simulator's, not measurements. The objects are simple shapes with plain colours, not scans of real products.

License and attribution

The dataset is released under CC-BY-NC-4.0. Please cite it as below. The code that writes, checks and scores it is Apache-2.0.

The renders show third-party robot models. No model file is distributed in this dataset.

• Robot. The UR7e arm from the Universal Robots ROS 2 description (BSD-3-Clause, Copyright Universal Robots A/S and ROS-Industrial contributors), and the Robotiq 2F-140 gripper from ros-industrial-attic/robotiq (BSD-2-Clause, Copyright 2013 ROS-Industrial). Neither Universal Robots, Robotiq nor ROS-Industrial endorses this dataset.
• Objects. Shapes and colours made by SpaceAI, released with this dataset under CC-BY-NC-4.0.

Citation

@misc{spaceai2026physgraspgen,
  title        = {PhysGraspGen soft-object grasps},
  author       = {{SpaceAI}},
  year         = {2026},
  version      = {1.0.0rc1},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.23787395},
  howpublished = {\url{https://huggingface.co/datasets/Official-Space-AI/physgraspgen-dataset}}
}

Contact

Questions and problems: GitHub issues or oss.spaceai@gmail.com.