geonmin-kim/so101_merged_v2

SO-101 merged v2 (1444 episodes, 486k frames)

Deduplicated merged dataset for SO-101 robot arm cube pick-and-place tasks. Combines five in-house recording pools plus whosricky/so101-megamix-v1.

Dataset composition

Source	Episodes	Status
geonmin-kim/SO101-large-small-cube-11dataset-ratio6231-0729	250+	Dedup: 183 unique
geonmin-kim/merged_all7	673	Dedup: 411 unique
geonmin-kim/merged_exp24_5cm_lv2_dagger_grasp_lv123	60	Dedup: 60 unique
geonmin-kim/merged_lv123_neartop	50	Dedup: 50 unique
geonmin-kim/SO101-lv2-single-cube-to-box-random-layout-v1	200	Dedup: 200 unique
Subtotal (in-house)	1233 nominal	904 unique
whosricky/so101-megamix-v1	400	400 (no dedup within megamix)
Total	1633	1444 unique

Deduplication method: Episode action array hash. Nominal count reflects recorded episodes; unique count removes duplicates within and across pools.

Statistics

{
  "total_episodes": 1444,
  "total_frames": 486351,
  "total_tasks": 33,
  "fps": 30,
  "codebase_version": "v3.0"
}

Frame duration: 486351 frames / 30 fps ≈ 270.8 hours.

Action space & normalization

Joint-space actions: 6-D (SO-101 arm DOF: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll + gripper).

Action statistics (so101_stats_stride1_v2.json, computed over 484,907 rows @ stride 1):

Dim	q01	q99
shoulder_pan dx (m)	-0.00924	+0.00960
shoulder_lift dy (m)	-0.01009	+0.00736
elbow_flex dz (m)	-0.01187	+0.00800
wrist_flex	—	—
wrist_roll	—	—
gripper	0.00000	0.55974

Important: Action quantiles differ from pre-megamix pools by ~12% on average (up to 19%). Megamix's slower teleop pulls the bounds inward. Stats v2 is not interchangeable with v1 — mixing them silently mis-scales actions.

Camera & resolution

• View: Ego (top-down, 30 fps)
• Resolution: 480×640 (height × width)
• Frames per episode: ≤ 332 (episode length varies; used stride 1)

Available cameras in source:

• observation.images.top — Both in-house and megamix
• observation.images.wrist — In-house pools only (not included in v2)

Tasks (33 unique)

Cube manipulation tasks, primarily:

• Pick up [color] cube
• Put [color] cube in box
• Stack cubes
• Rearrange cubes

Examples: "Pick up the large orange cube and put it in the box." (in English).

Use in Cosmos3-Super training

This dataset was used to post-train Cosmos3-Super for forward-dynamics world modeling:

from cosmos_framework.data.generator.action.datasets.so101_lerobot_dataset import SO101LeRobotDataset

ds = SO101LeRobotDataset(
    root="/path/to/so101_merged_v2",
    action_stats_path="/path/to/so101_stats_stride1_v2.json",
    camera="top",
    chunk_length=16,  # or 32 for horizon experiments
    mode="forward_dynamics",
)

Training recipe used:

• LoRA rank 16 on Cosmos3-Super (generation expert attention projections)
• Action pathway unfrozen (21.1M parameters)
• FSDP 4-way shard (4× A100 80GB)
• 30000 iterations @ selective activation checkpointing + torch.compile

Results: iter_000004000 checkpoint achieves ~0.5 s usable rollout horizon on held-out episodes.

Directory structure

so101_merged_v2/
├── data/
│   ├── chunk-{000..N}/
│   │   └── file-{0..M}.parquet       # Episode metadata (action, state, task)
│   └── ...
├── meta/
│   ├── info.json                     # Dataset summary
│   ├── tasks.parquet                 # Unique task list
│   ├── episodes/
│   │   └── chunk-{...}/file-{...}.parquet  # Episode-level info
│   └── ...
└── videos/
    └── observation.images.top/
        ├── chunk-{000..N}/
        │   └── file-{0..M}.mp4       # 30 fps, 480×640 RGB
        └── ...

Video files: 19 symlinks in original repo are realized as actual mp4s in this upload (3.6 GB total video).

Citation

If you use this dataset, please cite the source pools and megamix:

@dataset{geonmin2024so101lv2,
  title={SO101-lv2-single-cube-to-box-random-layout-v1},
  author={Geonmin Kim},
  year={2024},
  url={https://huggingface.co/datasets/geonmin-kim/SO101-lv2-single-cube-to-box-random-layout-v1}
}

@dataset{ricky2024so101megamix,
  title={so101-megamix-v1},
  author={Ricky Zheng},
  year={2024},
  url={https://huggingface.co/datasets/whosricky/so101-megamix-v1}
}

And the training work:

@software{xpucos3sim2024,
  title={xpu-cosmos3-simulator: Real-to-sim robot policy evaluator on Cosmos3},
  author={Geonmin Kim},
  year={2024},
  url={https://github.com/nota-github/xpu-cosmos3-simulator}
}

License

Apache License 2.0. Inherited from source pools (all Apache-compatible).

Notes

• Deduplication is deterministic (action hash-based) but not fully reproducible from first principles — source pool merge order and metadata differ across runs.
• Megamix integration: whosricky/so101-megamix-v1 is from a different lab with different teleop dynamics. Slower gripper and gentler motions pull action quantiles inward. Use so101_stats_stride1_v2.json for correct normalization.
• Wrist camera: Not included. Source pools contain observation.images.wrist but only top camera was used for Cosmos3 training. Wrist video can be extracted from original pools if needed.

---

Last updated: 2026-08-13
Dataset size: 3.9 GB (videos realized from symlinks)