serving_ab — Café Pick-and-Place (Long Episode + Atomic Demos)

LeRobot v3.0 dataset for OpenManipulator-X (OMX) café serving task. This is the merged dataset combining Category A (Long Episode) and Category B (Atomic Demos) used for the full GR00T N1.7 fine-tuning schedule.

---

Dataset summary

Item	Value
Robot	OpenManipulator-X 4-DOF (omx_follower)
Codebase version	v3.0
Total episodes	425 (A: 265, B: 160)
Total frames	219,877
Total duration	~2h 2m @ 30 fps
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)	74,679	34.0%
1 (cup)	96,803	44.0%
2 (home)	48,395	22.0%

---

Composition

Category A — Long Episode (265 episodes, eps 0~264)

Multi-step pick-and-place scenarios starting and ending at home pose.

sub-id	scenario	env state	total
A-1	one cup	cup=1	35
A-2	one plate	plate=1	35
A-3	two plates	plate=2	40
A-4	two cups	cup=2	40
A-5	one cup + one plate	plate=1, cup=1	50
A-6	two cups + one plate	plate=1, cup=2	45
A-7	empty (stop)	empty / leftovers	20

Category B — Atomic Demos (160 episodes, eps 265~424)

Skill-focused demonstrations to strengthen foundational grip reliability and multi-object handling.

sub-id	skill	total
B-1	plate single pick-and-place	50
B-2	cup single pick-and-place	50
B-3	sequential two plates	20
B-4	sequential 2~3 cups	20
B-5	retry recovery	20

Pick locations follow a 3×4 grid (12 positions, 7.5 cm spacing) spanning the OMX reachable workspace; each cell is visited 4~5 times in B-1/B-2 and 3 times in A-1/A-2 for position diversity.

---

Label boundary scheme

Multi-task episodes are segmented into per-action task ranges. Boundaries were auto-detected from gripper state events (fully_close → next start_open); missing boundaries in incomplete episodes were estimated from the median position percentile of complete episodes in the same scenario.

sub-id	segment pattern
A-1	cup(1) → home(2)
A-2	plate(0) → home(2)
A-3	plate(0) → plate(0) → home(2) (per-plate segment restart)
A-4	cup(1) → cup(1) → home(2) (per-cup segment restart)
A-5	cup(1) → plate(0) → home(2)
A-6	cup(1) → cup(1) → plate(0) → home(2)
A-7	home(2) only
B-1	plate(0) → home(2)
B-2	cup(1) → home(2)
B-3	plate(0) → plate(0) → home(2)
B-4	cup(1) → cup(1) → home(2) (or 3 cups)
B-5	<obj> → <obj> → home(2) (retry segments same task_index)

---

Repository layout

serving_ab/
├── data/chunk-000/file-000.parquet           # all 219,877 frames concatenated
├── meta/
│   ├── info.json                              # codebase_version=v3.0, totals
│   ├── tasks.parquet                          # task_index → instruction
│   ├── tasks.jsonl                            # mirror of tasks.parquet
│   ├── modality.json                          # GR00T modality mapping (state/action/video/annotation)
│   ├── stats.json                             # aggregate stats (mean/std/min/max/q01/q99)
│   └── episodes/chunk-000/file-000.parquet    # per-episode metadata + stats (107 cols)
└── videos/
    ├── observation.images.top/chunk-000/file-000.mp4    # ~924 MB
    └── observation.images.wrist/chunk-000/file-000.mp4  # ~1.59 GB

---

Companion datasets

Repo	Format	Episodes	Notes
jae0311/serving_a	v3.0	265	Category A only (long episodes)
jae0311/serving_b	v2.1	160	Category B only (atomic demos)

---

Recommended training schedule (GR00T N1.7)

Stage	Data	max-steps	LR	Goal
1	serving_b	5,000	1e-4	atomic pick-and-place reliability (90%+)
2	serving_ab	25,000	5e-5	long-episode performance from Stage 1 checkpoint

• effective batch size: 32 (per_device_batch=8 × grad_accum=4)
• action horizon: 16
• color jitter: brightness 0.2, contrast 0.2, saturation 0.4, hue 0.08

---

Acknowledgments

• Model: GR00T N1.7 (NVIDIA)
• Dataset format: LeRobot
• Robot: ROBOTIS OpenManipulator-X