File size: 6,065 Bytes
645e64f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
---
license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - manipulation
  - pick-and-place
  - cafe-serving
  - omx
size_categories:
  - 100K<n<1M
---

# serving_a — Café Pick-and-Place (Long Episode)

LeRobot **v3.0** dataset for OpenManipulator-X (OMX) cafe serving task.
This is the **Category A (Long Episode)** subset of the GR00T N1.7 fine-tuning dataset.

A converted **v2.1** version (per-episode parquet/mp4 layout, GR00T-trainable) is produced from this dataset using `scripts/lerobot_conversion/convert_v3_to_v2.py`.

---

## Dataset summary

| Item               | Value                                          |
| ------------------ | ---------------------------------------------- |
| Robot              | OpenManipulator-X 4-DOF (`omx_follower`)       |
| Codebase version   | `v3.0`                                         |
| Total episodes     | 265                                            |
| Total frames       | 144,788                                        |
| FPS                | 30                                             |
| Cameras            | `observation.images.top`, `observation.images.wrist` (640×480, 30 fps) |
| State / Action dim | 6 each (single_arm × 5, gripper × 1)           |
| Action horizon     | 16 (configured via `examples/OMX/omx_config.py`) |
| Tasks              | 3 (see below)                                  |

## Tasks (`task_index`)

| index | instruction                                |
| :---: | ------------------------------------------ |
| 0     | `pick up plate and place at target zone`   |
| 1     | `pick up cup and place at target zone`     |
| 2     | `return to home`                           |

### task distribution

| task_index | frames | pct  |
| ---------- | -----: | ---: |
| 0 (plate)  | 49,698 | 34.3% |
| 1 (cup)    | 66,874 | 46.2% |
| 2 (home)   | 28,216 | 19.5% |

---

## Scenario composition (Category A — Long Episode)

Every episode starts at home pose, performs the task(s), and returns to home.
Each scenario was collected as a separate sub-dataset (`serving_a1``serving_a7`) and then merged.

| sub-id | scenario               | env state            | success | retry | total |
| :----: | ---------------------- | -------------------- | ------: | ----: | ----: |
| A-1    | one cup                | cup=1                | 30      | 5     | **35** |
| A-2    | one plate              | plate=1              | 30      | 5     | **35** |
| A-3    | two plates             | plate=2              | 30      | 10    | **40** |
| A-4    | two cups               | cup=2                | 30      | 10    | **40** |
| A-5    | one plate + one cup    | plate=1, cup=1       | 38      | 12    | **50** |
| A-6    | one plate + two cups   | plate=1, cup=2       | 35      | 10    | **45** |
| A-7    | empty env (stop)       | empty (2 sub-cases)  | 18      |  0    | **18** |
|        | **Total**              |                      | **211** | **52**| **265**|

* **success**: completed normally.
* **retry**: gripper miss → short back-off → retry → success, all within a single episode.
* **A-7 sub-cases**: 7 episodes with both workspaces empty; 11 episodes with pick workspace empty and place workspace already containing previously-served objects (forces VLM to learn "stop when nothing to pick" even with clutter in the place zone).

---

## Label boundary scheme

Multi-task episodes are segmented into per-action task ranges so the policy can learn the language–action grounding.

| sub-id | segment pattern                                              | # boundaries |
| :----: | ------------------------------------------------------------ | :----------: |
| A-1    | `cup(1) → home(2)`                                           | 1            |
| A-2    | `plate(0) → home(2)`                                         | 1            |
| A-3    | `plate(0) → plate(0) → home(2)` (per-plate restart)          | 2            |
| A-4    | `cup(1) → cup(1) → home(2)` (per-cup restart)                | 2            |
| A-5    | `cup(1) → plate(0) → home(2)`                                | 2            |
| A-6    | `cup(1) → cup(1) → plate(0) → home(2)`                       | 3            |
| A-7    | `home(2)` only                                               | 0            |

Boundaries were auto-detected from gripper signal events (`fully_close` → next `start_open`) and audited manually for edge cases. Episodes that lacked the expected number of events had their missing boundary estimated from the median percentile of complete episodes within the same scenario.

---

## Repository layout

```
serving_a/
├── data/chunk-000/file-000.parquet           # all frames, concatenated
├── meta/
│   ├── info.json                              # codebase_version, features, totals
│   ├── tasks.parquet                          # task_index → instruction
│   └── episodes/chunk-000/file-000.parquet    # per-episode metadata + stats
└── videos/
    ├── observation.images.top/chunk-000/file-000.mp4
    └── observation.images.wrist/chunk-000/file-000.mp4
```

---

## Companion dataset

This is **Category A only**. The matching **Category B (atomic-focused demos)** dataset is available as:

> [`jae0311/serving_b`](https://huggingface.co/datasets/jae0311/serving_b) — 160 episodes, 75,089 frames

For GR00T N1.7 fine-tuning we use a 2-stage schedule:

| Stage | Data            | max-steps | LR     | Goal                                      |
| :---: | --------------- | :-------: | :----: | ----------------------------------------- |
| 1     | `serving_b` (B) | 5,000     | 1e-4   | atomic pick-and-place reliability         |
| 2     | A+B merged      | 25,000    | 5e-5   | full long-episode performance from Stage 1 ckpt |

---

## Citation / acknowledgments

- Model: [GR00T N1.7](https://huggingface.co/nvidia/GR00T-N1.7-3B) (NVIDIA)
- Dataset format: [LeRobot](https://github.com/huggingface/lerobot)
- Robot: ROBOTIS OpenManipulator-X