---
license: mit
task_categories:
  - robotics
tags:
  - lerobot
  - robotics
  - imitation-learning
  - manipulation
  - bimanual
  - ur3
size_categories:
  - 10K<n<100K
---

# EmbodyX UR3 — three bimanual manipulation tasks (LeRobot v2.1)

Real-world teleoperated demonstrations on a **bimanual dual-arm UR3**, released as the three
tasks used to fine-tune [`armanakbari4/imagewam-ur3-3task`](https://huggingface.co/armanakbari4/imagewam-ur3-3task).

| task | episodes | frames | fps | instruction |
|---|---|---|---|---|
| `blue_basket_lerobot` | 100 | 29,653 | 15 | "put the medicine then the measuring tape inside the blue basket" |
| `drawer_lerobot` | 100 | 32,982 | 15 | "open the drawer, put the white box inside the drawer then close the drawer" |
| `stacking_cubes_lerobot` | 100 | 37,042 | 15 | "put the green cube on top of the black cube and put the red cube on top of the green cube" |

**300 episodes / 99,677 frames** total, ~809 MB.

## Format

LeRobot **v2.1**, one directory per task, each with `data/`, `videos/`, and `meta/`.

**Observations** — three RGB cameras at **240×320**:

- `observation.images.camera_top`
- `observation.images.camera_wrist_left`
- `observation.images.camera_wrist_right`

**Action & state** — 14-dim each, absolute joint space, laid out as
`[L_arm(6) | L_gripper | R_arm(6) | R_gripper]`. Grippers are binary (0 = open, 1 = closed).

```python
from datasets import load_dataset
# or use the LeRobot loader against a local clone:
#   huggingface-cli download armanakbari4/ur3-3task-lerobot --repo-type dataset --local-dir ur3
```

## Known issue: declared codec is wrong for one task

`meta/info.json` declares `video.codec: av1` for all three tasks. That is **correct for
`drawer_lerobot` and `stacking_cubes_lerobot`, but wrong for `blue_basket_lerobot`**, whose
videos were re-encoded to **h264** without the metadata being updated. Verified with `ffprobe`:

| task | declared | actual |
|---|---|---|
| `blue_basket_lerobot` | av1 | **h264** |
| `drawer_lerobot` | av1 | av1 |
| `stacking_cubes_lerobot` | av1 | av1 |

Decoding is driven by the container rather than `info.json`, so this is harmless in practice —
but don't go looking for an AV1 decoder problem when reading `blue_basket`.

## Decoding note (AV1)

Two of the three tasks are genuinely AV1. If your `torchcodec` install fails to load, LeRobot
falls back to a pyav decoder that can leak memory heavily on these files and OOM dataloader
workers. On a CUDA-11 torch build this usually means the NPP libraries are missing
(`libnppicc.so.11`); installing `nvidia-npp-cu11` and putting it on `LD_LIBRARY_PATH` fixes it.

## Provenance

Collected by **EmbodyX** on a bimanual UR3 setup. Released here as the exact subset used for
the ImageWAM fine-tune above.