GraspTwin benchmark scenes

Digital twins of real tabletop scenes for language-conditioned grasping: one reconstructed .glb mesh per object, the object poses needed to rebuild each scene in simulation, and the natural-language tasks with their target objects.

• 9 scenes, 5 subscenes each (45 subscenes total)
• 8 household objects per subscene (scene_8_4 has 10)
• 10 natural-language tasks per scene, each tied to a target object
• about 2.6 GB of meshes

Layout

scene_N/
  tasks.json                     # 10 task strings, e.g. "grab the mug by the handle"
  scene_N_M/                     # subscene M = 1..5 (same objects, different arrangement)
    objects.json                 # names of the objects in the subscene
    target_objects.json          # target object per task, index-aligned with ../tasks.json
    boxes.json                   # 2D box [x1, y1, x2, y2] in pixels of the 1920x1080
                                 #   scene capture, per object name
    meshes/
      <object name>.glb          # one mesh per object (trimesh export)
      transforms.json            # per-object translation, rotation_wxyz, scale
      settle_transforms.json     # same, after the objects settle under physics in sim

Object names are the keys used across objects.json, boxes.json, transforms.json and the mesh filenames, so a task can be resolved to its mesh by name.

Loading a subscene

import json, trimesh
from pathlib import Path

sub = Path("scene_1/scene_1_1")
tasks = json.loads((sub.parent / "tasks.json").read_text())
targets = json.loads((sub / "target_objects.json").read_text())
poses = json.loads((sub / "meshes" / "transforms.json").read_text())["objects"]

for task, obj in zip(tasks, targets):
    mesh = trimesh.load(sub / "meshes" / f"{obj}.glb")
    print(task, "->", obj, poses[obj]["translation"])

License

MIT.