Add task_instruction.json files + dataset card
Browse files- README.md +25 -12
- hide_block_with_cup/task_instruction.json +40 -0
README.md
CHANGED
|
@@ -13,25 +13,34 @@ size_categories:
|
|
| 13 |
|
| 14 |
# Real Robot Manipulation Dataset
|
| 15 |
|
| 16 |
-
|
| 17 |
-
gripper
|
| 18 |
-
|
|
|
|
| 19 |
|
| 20 |
## Tasks
|
| 21 |
|
| 22 |
-
| Folder | Instruction |
|
| 23 |
-
|---|---|
|
| 24 |
-
| `put_blocks_in_plate` | Pick two blocks from the table and place them on the plate. |
|
| 25 |
-
| `remove_blocks_from_plate` | Pick two blocks off the plate and place them on the table. |
|
| 26 |
-
| `hide_blocks_with_bowl` | Pair the two blocks together and cover them with the bowl. |
|
| 27 |
-
| `remove_blocks_under_bowl` | Uncover the hidden blocks and spread the three objects out. |
|
| 28 |
-
| `hide_blocks_with_cup` | Cover each block with an upside-down cup. |
|
| 29 |
-
| `put_blocks_on_cup` | Lift each cup off a block and stack the block on top of the cup. |
|
| 30 |
-
| `put_blocks_down` | Take the blocks off the cups and place them on the table. |
|
|
|
|
| 31 |
|
| 32 |
Each task folder includes a `task_instruction.json` with the full task spec
|
| 33 |
(objects, pick/place strategies, gripper modes, workspace bounds).
|
| 34 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 35 |
## Layout
|
| 36 |
|
| 37 |
```
|
|
@@ -50,11 +59,15 @@ Each task folder includes a `task_instruction.json` with the full task spec
|
|
| 50 |
images/
|
| 51 |
camera1 (T,H,W,3) uint8 # base, 640x480 RGB
|
| 52 |
camera2 (T,H,W,3) uint8 # wrist, 640x480 RGB
|
|
|
|
| 53 |
qpos (T,7) float32 # 6 joints + gripper (norm)
|
| 54 |
state (T,14) float32 # tcp_pos(3)+tcp_quat(4)+joints(6)+gripper(1)
|
| 55 |
tcp_pos, tcp_quat, joints, gripper_pos
|
| 56 |
```
|
| 57 |
|
|
|
|
|
|
|
|
|
|
| 58 |
## meta.json schema (per episode)
|
| 59 |
|
| 60 |
For all tasks the meta contains the per-episode ground-truth object positions:
|
|
|
|
| 13 |
|
| 14 |
# Real Robot Manipulation Dataset
|
| 15 |
|
| 16 |
+
Tabletop manipulation tasks recorded on a **UR5e** with a **Robotiq 2F-140**
|
| 17 |
+
gripper. Most tasks use two **Intel RealSense** cameras (fixed base + wrist-mounted);
|
| 18 |
+
the newest `hide_block_with_cup` task adds a third (opposite-side counter view).
|
| 19 |
+
All recordings at **20 Hz**.
|
| 20 |
|
| 21 |
## Tasks
|
| 22 |
|
| 23 |
+
| Folder | Eps | Cameras | Instruction |
|
| 24 |
+
|---|---|---|---|
|
| 25 |
+
| `put_blocks_in_plate` | 100 | 2 | Pick two blocks from the table and place them on the plate. |
|
| 26 |
+
| `remove_blocks_from_plate` | 100 | 2 | Pick two blocks off the plate and place them on the table. |
|
| 27 |
+
| `hide_blocks_with_bowl` | 100 | 2 | Pair the two blocks together and cover them with the bowl. |
|
| 28 |
+
| `remove_blocks_under_bowl` | 100 | 2 | Uncover the hidden blocks and spread the three objects out. |
|
| 29 |
+
| `hide_blocks_with_cup` | 100 | 2 | Cover each block with an upside-down cup. |
|
| 30 |
+
| `put_blocks_on_cup` | 100 | 2 | Lift each cup off a block and stack the block on top of the cup. |
|
| 31 |
+
| `put_blocks_down` | 100 | 2 | Take the blocks off the cups and place them on the table. |
|
| 32 |
+
| `hide_block_with_cup` | 120 | **3** | Cover the block with the upside-down cup. *(eps 1-66 light, 67-120 dark)* |
|
| 33 |
|
| 34 |
Each task folder includes a `task_instruction.json` with the full task spec
|
| 35 |
(objects, pick/place strategies, gripper modes, workspace bounds).
|
| 36 |
|
| 37 |
+
**`hide_block_with_cup` specifics**: single-task variant (1 cup + 1 block).
|
| 38 |
+
Recording captures only the cup-pick + cup-place motion; the
|
| 39 |
+
inter-episode object randomization is **not** recorded. A 3rd camera
|
| 40 |
+
(`camera3`, oppo-countertop) was added for this task. Per-episode `meta.json`
|
| 41 |
+
includes a `lighting` field (`"light"` or `"dark"`) — the room lighting
|
| 42 |
+
changed at episode 67.
|
| 43 |
+
|
| 44 |
## Layout
|
| 45 |
|
| 46 |
```
|
|
|
|
| 59 |
images/
|
| 60 |
camera1 (T,H,W,3) uint8 # base, 640x480 RGB
|
| 61 |
camera2 (T,H,W,3) uint8 # wrist, 640x480 RGB
|
| 62 |
+
camera3 (T,H,W,3) uint8 # (hide_block_with_cup only) oppo-countertop
|
| 63 |
qpos (T,7) float32 # 6 joints + gripper (norm)
|
| 64 |
state (T,14) float32 # tcp_pos(3)+tcp_quat(4)+joints(6)+gripper(1)
|
| 65 |
tcp_pos, tcp_quat, joints, gripper_pos
|
| 66 |
```
|
| 67 |
|
| 68 |
+
The HDF5 root attribute `num_cameras` reports the actual count for that file
|
| 69 |
+
(2 for all 7 original tasks, 3 for `hide_block_with_cup`).
|
| 70 |
+
|
| 71 |
## meta.json schema (per episode)
|
| 72 |
|
| 73 |
For all tasks the meta contains the per-episode ground-truth object positions:
|
hide_block_with_cup/task_instruction.json
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"robot": "UR5e + Robotiq 2F-140 gripper",
|
| 3 |
+
"cameras": {
|
| 4 |
+
"camera1": "fixed base RealSense (table-facing)",
|
| 5 |
+
"camera2": "wrist-mounted RealSense",
|
| 6 |
+
"camera3": "oppo-countertop RealSense (opposite-side counter view)"
|
| 7 |
+
},
|
| 8 |
+
"camera_resolution": [
|
| 9 |
+
640,
|
| 10 |
+
480
|
| 11 |
+
],
|
| 12 |
+
"control_hz": 20,
|
| 13 |
+
"tcp_orientation_axis_angle": [
|
| 14 |
+
3.14159,
|
| 15 |
+
0.0,
|
| 16 |
+
0.0
|
| 17 |
+
],
|
| 18 |
+
"workspace_trapezoid": {
|
| 19 |
+
"y_range": [
|
| 20 |
+
0.5,
|
| 21 |
+
0.8
|
| 22 |
+
],
|
| 23 |
+
"x_halfwidth_at_y_min": 0.1,
|
| 24 |
+
"x_halfwidth_at_y_max": 0.18
|
| 25 |
+
},
|
| 26 |
+
"episodes": 120,
|
| 27 |
+
"task": "hide_block_with_cup",
|
| 28 |
+
"instruction": "Cover the block with the upside-down cup.",
|
| 29 |
+
"description": "One cup (upside down, rim on table; R = 8 cm) and one block start at random separated positions (>= 20 cm apart). The robot picks the cup (fully open gripper) and places it over the block via moveUntilContact so the rim lands on the table around the block, hiding it. The recording ends after the cup is placed. Each episode also has an unrecorded RESET phase that lifts the cup off, places it at a new random position, then moves the block to its own new random position -- this produces the random starting pose for the next episode.",
|
| 30 |
+
"objects": [
|
| 31 |
+
"cup (R=8cm, upside down)",
|
| 32 |
+
"block (~2.5cm cube)"
|
| 33 |
+
],
|
| 34 |
+
"gripper_modes": "block-pick narrows to 50% open; cup-pick fully open",
|
| 35 |
+
"lighting_split": {
|
| 36 |
+
"light": "episodes 1-66",
|
| 37 |
+
"dark": "episodes 67-120",
|
| 38 |
+
"note": "Room lighting changed mid-collection. Each episode's meta.json has a `lighting` field with value 'light' or 'dark'."
|
| 39 |
+
}
|
| 40 |
+
}
|