dexgrasp-robotwin-4gripper

Physics-verified grasp annotations for 788 RoboTwin / RoboDojo manipulation assets, each annotated with four grippers: dex1_1 (Unitree Dex1-1), piper (AgileX Piper), sawyer_umi (Sawyer + UMI), stretch3 (Hello Robot Stretch 3).

Companion to prlq/dexgrasp-omni-4gripper-v2 (OmniObject3D assets, same protocol). An earlier release of these same objects, prlq/dexgrasp-2gripper, used a different protocol and is left unchanged — see Relationship to the earlier release.

Layout

<object>/
  Object.usd
  Annotation/
    dex1_1_grasp_pose.json
    piper_grasp_pose.json
    sawyer_umi_grasp_pose.json
    stretch3_grasp_pose.json
manifest.csv              per-object, per-gripper candidates / passed / pass rate,
                          plus the category and the density actually used
stats_by_gripper.csv      aggregate pass rate and object usability per gripper
stats_by_category.csv     per-category pass rate and zero-pass count per gripper
stats_by_density.csv      pass rate per density bucket per gripper

Each *_grasp_pose.json holds the grasps that survived physics replay, plus:

"evaluation": {"collider": "coacd", "num_candidates": 512, "num_passed": 4}

num_passed equals the length of functional_grasp.body. Every pose was simulated, squeezed, and shaken under gravity without dropping the object.

How it was produced

1. Synthesis — BODex/cuRobo force-closure synthesis, 2048 seeds per (object, gripper), best-first ordering. (Synthesis does not use object mass: its config sets enable_density: False, so the density table below affects only step 2.)
2. Physics replay — Isaac Sim 5.1 headless, the first 512 candidates of each pair, 256 parallel environments (64 on smaller cards). The hand approaches from 5 cm behind the annotated pregrasp, closes, then gravity is enabled. Colliders are the CoACD convex decomposition the candidates were synthesised against.

setting	value
object density	per-category table, 246 entries (metal 4000, ceramic/glass 2200, hard plastic 1300, wood 700, hollow container 400, else 1000)
mass	density x signed mesh volume
contact friction	static 0.5 / dynamic 0.5, combined by average
restitution	0.0
candidates replayed	first 512 of 2048

Category keys are normalised before density lookup: a leading <digits>_, a trailing _<digits>, and the rigid_/clutter_/geometry_ source-set prefixes are stripped. Those prefixes carry no material information, and leaving them in dropped table coverage to 8%. The per-object density actually used is in manifest.csv.

Pass rates

On the 788 published objects (stats_by_gripper.csv):

gripper	candidate pass rate	zero-pass objects	object usability
stretch3	33.88%	0	100%
dex1_1	26.41%	0	100%
piper	25.51%	0	100%
sawyer_umi	1.11%	277	64.9%

The three leading grippers show zero zero-pass objects by construction — that is the filter, not a property of the data. Before filtering, over all 1113 evaluated objects, the rates were stretch3 24.66%, dex1_1 19.07%, piper 18.29%, sawyer_umi 0.85%.

🔴 Gripper ranking is dataset-dependent

dataset	ranking
this one (RoboTwin / RoboDojo props)	stretch3 > dex1_1 > piper
OmniObject3D	dex1_1 > stretch3 > piper

Identical physics protocol in both, so the difference comes from object composition: this set is heavy on blocks, number tiles and mahjong pieces, where stretch3 has the advantage. Do not conclude which gripper is better from one dataset. Overall rates are also lower here than on OmniObject3D (25–34% vs 37–41%) for the same reason — manipulation props are harder to grasp than the food items that dominate OmniObject3D.

🔴 Noise floor: about 0.6%

Measured on the OmniObject3D companion set by replaying objects whose density was identical in both passes, i.e. byte-identical physics inputs: the pass rate still moved 0.6%, from GPU PhysX non-determinism plus differing environment counts (64 vs 256). Treat any difference under about 1% as noise.

What was filtered out

step	removed	left
objects evaluated with all four grippers	-	1113
zero surviving grasps for any of dex1_1 / piper / stretch3	325	788

sawyer_umi is deliberately excluded from the filter: its aperture ceiling is a property of the gripper, and gating on it would have removed most of the dataset. It therefore still has 277 objects with an empty annotation here, and is pending a redesign.

Relationship to the earlier release

prlq/dexgrasp-2gripper covers the same assets but is not comparable to this one:

	earlier release	this release
grippers	piper, dex1_1	all four
candidates replayed	all (~2036)	first 512
density	uniform 1000	per-category table
objects	1113	788 (filtered)

Pass rates differ mostly because of the denominator: replaying all 2036 candidates includes the low-quality tail of a quality-ordered list, so its rates look lower (dex1_1 11.54%, piper 7.46%) without the annotations being worse. The candidate list is quality-ordered, so a prefix of it is not a random sample and rates across different candidate counts should not be compared.

Known limitations

• No textures. These assets ship as geometry only; there is no textures/ directory. This does not affect the annotations: physics uses the collider, the mass and a global friction material, never a visual material.
• CoACD collider pieces are not included. Re-verifying a grasp requires regenerating the decomposition from Object.usd; the result will be close but not bit-identical. Ask if you need the exact pieces.
• Density is a material estimate, not a measurement. Categories were assigned to buckets by material keyword; manifest.csv records what each object actually used so an independent choice can be substituted.

Source data

Objects come from the RoboTwin and RoboDojo asset sets; respect their upstream licenses for the meshes. The annotations under Annotation/ and all CSVs are released under CC-BY-4.0.