---
license: apache-2.0
task_categories:
- robotics
tags:
- robotics
- simulation
- isaac-sim
- franka
- discodemo
pretty_name: DiscoDemo raw states — StackCube
---

# DiscoDemo raw states: StackCube

Camera-free simulator state trajectories of the generated demonstrations for the **StackCube** task
("stack the red cube on the blue cube") on a real-to-sim Franka FR3 workcell in Isaac Sim, from
[DiscoDemo](https://davian-robotics.github.io/DiscoDemo/). Each file holds the 3,000 successful, safety-filtered
trajectories of one demonstration generator. Use them to analyze or re-render the demonstrations without
running the generators (object and end-effector trajectories, diversity and coverage analyses).

These are the trajectories of the generated datasets, in the same order: `traj_<i>` of `discodemo.h5` is episode `i`
of [`DAVIAN-Robotics/DiscoDemo-Stage2_GenData-StackCube`](https://huggingface.co/datasets/DAVIAN-Robotics/DiscoDemo-Stage2_GenData-StackCube),
and `traj_<i>` of `prfcl.h5` is episode `i` of
[`DAVIAN-Robotics/DiscoDemo-Stage2_GenData-StackCube-alpha0`](https://huggingface.co/datasets/DAVIAN-Robotics/DiscoDemo-Stage2_GenData-StackCube-alpha0).

| File | Generator | Trajectories | Steps (median / max) |
|---|---|---|---|
| `discodemo.h5` | DiscoDemo generator (diversity weight alpha = 0.3) | 3,000 | 163 / 598 |
| `prfcl.h5` | P-RFCL baseline (alpha = 0) | 3,000 | 155 / 259 |

## Format

One HDF5 file per generator, trajectories `traj_0` ... `traj_2999`, recorded at 20 Hz:

```
traj_<i>/
  actions                    (T, 8)    absolute FR3 joint position targets for joints 1-7 in the DROID
                                       frame (joint 7 offset by -pi/4) and the binary gripper command (1 = close)
  success                    (T+1,)    task success predicate at each state
  states/
    robot/joint_pos          (T+1, 9)  7 arm joints + 2 finger joints
    robot/joint_vel          (T+1, 9)
    objects/
      cube_blue/root_pose  (T+1, 7)   position (env frame) + quaternion (w, x, y, z)
      cube_blue/root_vel   (T+1, 6)   linear + angular velocity (world frame)
      cube_red/root_pose  (T+1, 7)   position (env frame) + quaternion (w, x, y, z)
      cube_red/root_vel   (T+1, 6)   linear + angular velocity (world frame)
  skill_z                    (3,)      skill latent sampled for the episode (discodemo.h5 only)
```

The root attributes give the language instruction, the task name and the recording rate. End-effector poses can be recovered from `joint_pos` with forward kinematics of the FR3 model in
the code repository.

## Related resources

- Code (environments, generators, data pipeline, analysis): https://github.com/DAVIAN-Robotics/DiscoDemo
- Models and datasets: https://huggingface.co/collections/DAVIAN-Robotics/discodemo-6ac696e70d028fddf0c3dfeb

## Citation

```bibtex
@article{park2026discodemo,
  title   = {DiscoDemo: Discovering Efficient and Diverse Robot Demonstrations for Imitation Learning},
  author  = {Park, Minho and Kim, Kinam and Kim, Donghu and Lee, Byungkun and Hwang, Dongyoon and Shin, Yongjae and Hyung, Junha and Lee, Hojoon and Choo, Jaegul},
  journal = {arXiv preprint},
  year    = {2026}
}
```