README.md

---
license: apache-2.0
task_categories:
  - robotics
tags:
  - real-robot
  - manipulation
  - ur5e
  - imitation-learning
size_categories:
  - 100K<n<1M
---

# Real Robot Manipulation Dataset

Tabletop manipulation tasks recorded on a **UR5e** with a **Robotiq 2F-140**
gripper. Most tasks use two **Intel RealSense** cameras (fixed base + wrist-mounted);
the newest `hide_block_with_cup` task adds a third (opposite-side counter view).
All recordings at **20 Hz**.

## Tasks

| Folder | Eps | Cameras | Instruction |
|---|---|---|---|
| `put_blocks_in_plate` | 100 | 2 | Pick two blocks from the table and place them on the plate. |
| `remove_blocks_from_plate` | 100 | 2 | Pick two blocks off the plate and place them on the table. |
| `hide_blocks_with_bowl` | 100 | 2 | Pair the two blocks together and cover them with the bowl. |
| `remove_blocks_under_bowl` | 100 | 2 | Uncover the hidden blocks and spread the three objects out. |
| `hide_blocks_with_cup` | 100 | 2 | Cover each block with an upside-down cup. |
| `put_blocks_on_cup` | 100 | 2 | Lift each cup off a block and stack the block on top of the cup. |
| `put_blocks_down` | 100 | 2 | Take the blocks off the cups and place them on the table. |
| `hide_block_with_cup` | 120 | **3** | Cover the block with the upside-down cup. *(eps 1-66 light, 67-120 dark)* |

Each task folder includes a `task_instruction.json` with the full task spec
(objects, pick/place strategies, gripper modes, workspace bounds).

**`hide_block_with_cup` specifics**: single-task variant (1 cup + 1 block).
Recording captures only the cup-pick + cup-place motion; the
inter-episode object randomization is **not** recorded. A 3rd camera
(`camera3`, oppo-countertop) was added for this task. Per-episode `meta.json`
includes a `lighting` field (`"light"` or `"dark"`) — the room lighting
changed at episode 67.

## Layout

```
<task>/
    1.hdf5  ...  100.hdf5     # per-episode trajectories (images + state)
    meta/
        1/meta.json  ...  100/meta.json   # per-episode ground truth
    task_instruction.json     # task spec
```

## HDF5 schema (per episode)

```
/timestamps                       (T,)      float64
/observations/
    images/
        camera1                   (T,H,W,3) uint8   # base, 640x480 RGB
        camera2                   (T,H,W,3) uint8   # wrist, 640x480 RGB
        camera3                   (T,H,W,3) uint8   # (hide_block_with_cup only) oppo-countertop
    qpos                          (T,7)     float32 # 6 joints + gripper (norm)
    state                         (T,14)    float32 # tcp_pos(3)+tcp_quat(4)+joints(6)+gripper(1)
    tcp_pos, tcp_quat, joints, gripper_pos
```

The HDF5 root attribute `num_cameras` reports the actual count for that file
(2 for all 7 original tasks, 3 for `hide_block_with_cup`).

## meta.json schema (per episode)

For all tasks the meta contains the per-episode ground-truth object positions:

```json
{
    "task": "<task name>",
    "episode": <int>,
    "start": {<objects>: <(x,y)>, ...},
    "end":   {<objects>: <(x,y)>, ...},
    "init_pose": [x, y, z]
}
```

Task-specific keys are additionally recorded (e.g. `perm` and `cup_order` for
the cup tasks, `sign` and `active`/`passive` for the bowl hide task,
`block_order` for the bowl/cup remove tasks).

## Robot setup

- **Robot**: UR5e via RTDE
- **Gripper**: Robotiq 2F-140 (auto-calibrated)
- **TCP orientation**: pointing straight down (axis-angle `[π, 0, 0]`)
- **Cameras**: 2× Intel RealSense, 640×480 @ 20 Hz
- **Workspace**: trapezoid with `y ∈ [0.50, 0.80]` and `x` half-width
  `0.10 → 0.18` (linearly with y)

## Per-task episode counts

| Task | Episodes | Frames (~) | Wall-clock (~) |
|---|---|---|---|
| put_blocks_in_plate | 100 | 45,576 | 38 min |
| remove_blocks_from_plate | 100 | 45,929 | 38 min |
| hide_blocks_with_bowl | 100 | 43,881 | 37 min |
| remove_blocks_under_bowl | 100 | 66,090 | 55 min |
| hide_blocks_with_cup | 100 | (~varies) | ~40 min |
| put_blocks_on_cup | 100 | (~varies) | ~50 min |
| put_blocks_down | 100 | (~varies) | ~30 min |

## License

Apache 2.0.