Official-Space-AI/physgraspgen-dataset

PhysGraspGen soft-object grasps 1.0.0

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	2,749,581
Episodes	6,733, successful grasps only
Object types	14
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 1.0.0 (Apache-2.0)
License	CC-BY-NC-4.0
DOI	10.5281/zenodo.24069711

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 6,733 episodes, 95% intervals):

Criterion	Lifted	Held	Placed	Task score
C, 5 cm for 2 s	6,733/6,733, 1.000 [0.999, 1.000]	6,733/6,733, 1.000 [0.999, 1.000]	6,733/6,733, 1.000 [0.999, 1.000]	1.000
B, 12 cm for 4 s	2,593/6,733, 0.385 [0.374, 0.397]	2,561/6,733, 0.380 [0.369, 0.392]	2,561/6,733, 0.380 [0.369, 0.392]	0.382

This release holds successful grasps only, so every episode meets C; the scores show what the episodes are, not how hard the task is. B asks for a lift of 12 cm held for 4 s, which the episodes were not selected for.

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	Shape and colour
apple	563	238,915	apple, about 80 mm	scanned by SpaceAI
banana	96	40,730	banana, about 155 mm long	scanned by SpaceAI
bike_grip	128	54,493	bicycle grip, about 105 mm long and 31 mm across	scanned by SpaceAI
eraser	743	314,610	whiteboard eraser, about 135 x 57 x 33 mm	scanned by SpaceAI
hose_connector	629	268,436	39 mm barbed tube connector	built by SpaceAI from its dimensions
lemon	744	316,176	lemon, about 65 mm	YCB Object and Model Set (CC BY 4.0)
orange	730	309,884	orange, about 72 mm	YCB Object and Model Set (CC BY 4.0)
pear	687	291,735	pear, about 95 mm tall	YCB Object and Model Set (CC BY 4.0)
puff_pastry	423	159,531	80 x 50 x 30 mm block, cream	built by SpaceAI from its dimensions
rubber_gasket	273	102,887	60 mm ring, black	built by SpaceAI from its dimensions
sausage	106	45,026	100 mm sausage of 15 g	built by SpaceAI from its dimensions
silicone_pad	407	153,436	80 x 40 x 25 mm pad, blue	built by SpaceAI from its dimensions
sponge	444	167,293	150 x 90 x 30 mm block, yellow	built by SpaceAI from its dimensions
tofu	760	286,429	85 x 110 x 45 mm packaged block, cream with a pale green lid	built by SpaceAI from its dimensions

The dataset is updated regularly: later releases add episodes and new object types.

Changes from 1.0.0rc1

• Nine object types added: apple, banana, bike_grip, eraser, hose_connector, lemon, orange, pear and sausage. They were generated in a second model of the cell, with its own table, camera placement and gripper pads; each episode's header.camera and facts.camera record its camera.
• Successful grasps only. 1.0.0rc1 also kept 179 failures (177 of them rubber_gasket), which this release leaves out; the successful episodes of the five earlier types are those of 1.0.0rc1, exported again under this version.

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. Each object's shape and colour come from the source the Object types table names. 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. Scanned shapes are of single specimens, and the material of every object is a model, not a measurement of that object.

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 and objects. 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 by SpaceAI. The shapes built from dimensions, their colours, and the scans of the apple, banana, bike grip and eraser are SpaceAI's, released with this dataset under CC-BY-NC-4.0.
• Objects from YCB. The lemon (014_lemon), pear (016_pear) and orange (017_orange) are derived from the YCB Object and Model Set (google_16k meshes and textures), licensed under CC BY 4.0. SpaceAI re-centred and uniformly rescaled each mesh, tetrahedralized it with fTetWild, gave it a material model, and downsampled its texture from 4096 to 2048 pixels; the frames show the simulated objects. The YCB authors do not endorse this dataset. B. Calli, A. Walsman, A. Singh, S. Srinivasa, P. Abbeel and A. M. Dollar, "Benchmarking in Manipulation Research: Using the Yale-CMU-Berkeley Object and Model Set", IEEE Robotics and Automation Magazine 22(3), 2015.

Citation

@misc{spaceai2026physgraspgen,
  title        = {PhysGraspGen soft-object grasps},
  author       = {{SpaceAI}},
  year         = {2026},
  version      = {1.0.0},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.24069711},
  howpublished = {\url{https://huggingface.co/datasets/Official-Space-AI/physgraspgen-dataset}}
}

Contact

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