---
license: apache-2.0
task_categories:
- robotics
size_categories:
- 10K<n<100K
---

# RealRobot push-blocks 🦾

Real-world Franka FR3 cube-pushing demonstrations with language instructions,
collected by replaying simulated expert trajectories open-loop on hardware
rather than by teleoperation. Each real recording is paired with the simulated
trajectory that produced it, so the dataset also serves as a measurement of the
sim-to-real gap.

📝 *A Sim-to-Real Integration Pipeline for Training and Deployment of
Chunk-Based VLA Manipulation Policies* (ISIR, Sorbonne Université / CNRS).
Pipeline source: <https://gitlab.isir.upmc.fr/kappel/sim2real_public_chunk_control/>

## At a glance 👁️

Example : `Push the red block left`
![Example](assets/teaser.png)

| | |
|---|---|
| Episodes | 200 (50 per task) |
| Frames | 11,267 |
| Episode length | 49–64 frames |
| Robot | Franka FR3, single arm |
| Camera | Orbbec Femto Bolt RGB, fixed, native 1000×1000 |
| Format | LeRobot v3.0 |

## Tasks

Four language-conditioned combinations of two motions and two target objects:

- `Push the red block left`
- `Push the red block right`
- `Push the blue block left`
- `Push the blue block right`

The instruction string is stored per frame as the LeRobot `task` field.

## Data fields

### `observation.state` — shape `(6,)`, float32

Absolute end-effector pose in the **robot base frame**, which is also the world
frame origin in both the simulated and real setups.

| index | name | meaning |
|---|---|---|
| 0–2 | `x, y, z` | end-effector position, metres |
| 3–5 | `rx, ry, rz` | end-effector orientation|

### `action` — shape `(7,)`, float32

End-effector **deltas**, not absolute poses. Applied cumulatively to the current
pose to obtain the next target frame.

| index | name | meaning |
|---|---|---|
| 0–2 | `dx, dy, dz` | translation increment, metres |
| 3–5 | `drx, dry, drz` | orientation increment, same convention as state |
| 6 | `gripper` | binary command|


### `observation.images.static` — `(H, W, 3)` uint8, H.264

Fixed third-person RGB view. Native capture is 1000×1000.

## Loading

```python
from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("clemgris/RealRobot_push-blocks")
print(ds.num_episodes, ds.num_frames, ds.fps)

sample = ds[0]
sample["observation.state"]           # (6,)
sample["action"]                      # (7,)
sample["observation.images.static"]   # (3, H, W)
sample["task"]                        # instruction string
```

## Citation

```bibtex
@misc{kappel2026sim2real,
  title  = {A Sim-to-Real Integration Pipeline for Training and Deployment
            of Chunk-Based VLA Manipulation Policies},
  author = {Kappel, Mathilde and Grislain, Cl\'emence and Chetouani, Mohamed
            and Sigaud, Olivier and Khoramshahi, Mahdi and Annabi, Louis
            and Ben Amar, Fa\"iz and Doncieux, St\'ephane},
  year   = {2026},
}
```