The Dataset Viewer has been disabled on this dataset.

Real2sim of ABC dataset

1,002 simulation environments for five bimanual manipulation tasks from the ABC dataset, each reconstructed from a real teleoperated episode by Versor. Read more in our blog post: Real-world data as the foundation for robotics simulations.

Real episodes from ABC (left in each pair) beside the scenes reconstructed from them (right): mug flipping, bottle placement, sorting screws and nuts, and packing snacks into a paper bag.

Each scene is a standalone USDZ with everything in place for policy training and evaluation: the robot cell, two YAM arms at the real episode's first joint angles, and the task's objects where the real episode had them, with mass, friction and collision shapes.

What's inside

Folder Task Scenes
put_plastic_bottles_in_bin/ Put the plastic bottles in the bin 200
load_plates_into_dish_rack/ Load the plates into the dish rack 198
turn_mug_right_side_up/ Turn the mugs right side up 272
sort_the_screws_and_nuts_into_containers/ Sort the screws and nuts into containers 132
place_the_snacks_into_the_paper_bag/ Place the snacks into the paper bag 200

Each folder holds <episode>.usdz (the scene), <episode>.json (what the scene holds and how it was checked) and index.json (all scenes in the folder). The episode id is the ABC-130k episode the scene was reconstructed from.

  • Units and frame: metres, kilograms, seconds; z up.
  • Robot: two fixed-base arms with six revolute joints and a two-finger gripper each, with joint position controllers initialized to the recorded episode’s starting commands.
  • Physics: rigid bodies with mass, inertia and friction; MuJoCo's contact and solver parameters are stored alongside in MuJoCo's USD schema.
  • Snacks: textured object models, and the head camera solved for each episode. The snack packets are deformable solids and the bag is a cloth, settled in Newton.

Quick start

Download one task, or drop --include for all of them (22 GB):

pip install -U huggingface_hub usd-core
hf download versor-tech/versor-abc-sim --repo-type dataset --include "turn_mug_right_side_up/*" --local-dir abc-sim

Open a scene with the USD Python API:

from pathlib import Path
from pxr import Usd, UsdPhysics

scene = next(Path("abc-sim/turn_mug_right_side_up").glob("*.usdz"))
stage = Usd.Stage.Open(str(scene))
bodies = [p.GetPath() for p in stage.Traverse() if p.HasAPI(UsdPhysics.RigidBodyAPI)]
print(scene.name, len(bodies), "rigid bodies")

The scenes also open directly in Isaac Sim. The snacks scenes need PhysX's GPU dynamics, which PhysX requires for deformables, or Newton.

License

Apache-2.0. The scenes build on ABC and ABC-130k, both Apache-2.0; see LICENSE and NOTICE.

Citation

If you use these scenes, please cite our post and the ABC paper:

@misc{versor2026abcsim,
  title        = {Real-world data as the foundation for robotics simulations},
  author       = {{Versor}},
  year         = {2026},
  howpublished = {\url{https://versor.tech/news/abc-sim/}}
}

@article{allshire2026abc,
  title   = {Scalable Behavior Cloning with Open Data, Training, and Evaluation},
  author  = {Arthur Allshire and Himanshu Gaurav Singh and Ritvik Singh and Adam Rashid and Hongsuk Choi and David McAllister and Justin Yu and Yiyuan Chen and Huang Huang and Pieter Abbeel and Xi Chen and Rocky Duan and Phillip Isola and Jitendra Malik and Fred Shentu and Guanya Shi and Philipp Wu and Angjoo Kanazawa},
  journal = {arXiv preprint arXiv:2606.27375},
  year    = {2026}
}
Downloads last month
17

Paper for versor-tech/versor-abc-sim