dexgrasp-2gripper

1113 household objects, each carrying physics-validated grasp annotations for two parallel-jaw grippers: AgileX Piper and Unitree Dex1-1. Every object in this dataset has both; nothing here is annotated for only one gripper.

Layout

<object>/
  Object.usd                               the asset, metres, Z up
  textures/                                its textures (when the source has them)
  Annotation/piper_grasp_pose.json         grasps that passed for Piper
  Annotation/dex1_1_grasp_pose.json        grasps that passed for Dex1-1
manifest.csv                               object, source, per-gripper candidate and pass counts

manifest.csv columns: object, source (robodojo_usd or robotwin_usd), piper_candidates, piper_passed, dex1_1_candidates, dex1_1_passed.

What an annotation file contains

{
  "type": "<object name>",
  "bottom_center": [x, y, z],        // object frame, metres: centre of its footprint
  "functional_grasp": {
    "body": [                        // one entry per grasp that PASSED the physics check
      {
        "coarse_grasp": {"position": [...], "orientation": [w, x, y, z], "joints": [...]},
        "fine_grasp":   {...},
        "final_grasp":  {...}
      }
    ]
  },
  "evaluation": {"collider": "coacd", "num_candidates": 2048, "num_passed": 137}
}

• Three stages per grasp. coarse_grasp is the pregrasp (hand open, standing off the object), fine_grasp is the pose where the jaws first close on it, and final_grasp is after squeezing. Replay them in that order.
• Frame. position/orientation are the pose of the gripper's palm link in the object frame — gripper_base for Piper, base_link for Dex1-1. Metres and a w-first quaternion.
• Joints. One value per driven joint, in the order below. Prismatic joints are in metres.
  • Piper: joint7, 0 shut → 0.035 open (70 mm between the pads). joint8 follows it with multiplier −1.
  • Dex1-1: Joint1_1, −0.02 is fully open (90 mm) and +0.0245 is shut; Joint2_1 follows with multiplier +1.
• Only passing grasps are stored. num_candidates is how many were tried, so num_passed / num_candidates is that object's pass rate for that gripper.

How the annotations were produced

1. Synthesis. BODex force-closure optimisation against the object's CoACD convex decomposition, 2048 seeds per object and gripper, friction 0.3. The output is an ordered list of candidate grasps — the front of the list is markedly better than the tail.
2. Physics validation. Every candidate is replayed in Isaac Sim 5.1 (PhysX GPU) in its own environment: approach from 5 cm back → close → squeeze → settle under gravity → lift 0.2 m → hold. A grasp passes only if the object never shifts more than 0.02 m or rotates more than 15° while the jaws close, does not drop, and is still in the hand at the end of the lift.
3. Objects are rigid, density 1000 kg/m³, collided as their CoACD pieces; object-to-pad friction 0.5/0.5 combined by averaging.

Every candidate was evaluated — this dataset is not a prefix of the candidate list. Later runs of the same pipeline may evaluate only the first N candidates for speed; those numbers are not comparable with these.

Caveats

• Pass rates vary by orders of magnitude between objects; flat, thin-walled scans are much harder than compact ones. Use manifest.csv to filter.
• The two grippers have different apertures (70 mm vs 90 mm), so an object that is ungraspable for Piper may still be annotated for Dex1-1 — with a lower pass count, never with zero: an object with no passing grasp for either gripper is still present if the synthesis produced candidates for both.
• Assets come from two upstream collections (robodojo_usd, robotwin_usd) and keep their original names; their licences are the upstream ones.
• Physics verdicts are not bit-exact across GPU architectures: a borderline grasp can flip. Measured on one object, 34 of 37 passing grasps were identical between runs, with 3 flipping each way.