Vi2Act

Vi2Act is a small synthetic video-to-action dataset for the Fourier GR1 humanoid. Videos are generated by a Cosmos-Predict2 14B Video2World model (GR00T-Dreams-GR1 checkpoint), then converted into LeRobot format.

Source pipeline: NVIDIA GR00T-Dreams / DreamGen using the EVAL-175 DreamGen Bench prompts and first-frame images.

What's in this repo

Three LeRobot subsets (5 episodes each):

Folder	Split (EVAL-175)	Episodes	Frames	FPS	Tasks
gr1_env_lerobot.data/	environment generalization	5	465	8	5
gr1_object_lerobot.data/	object generalization	5	465	8	5
gr1_behavior_lerobot.data/	behavior / skill generalization	5	465	8	5

Totals: 15 episodes, 1,395 frames, 15 language tasks, 1 ego-view camera.

This is a pilot / smoke-test dump (NUM_SAMPLES=5 per subset), not the full EVAL-175 benchmark.

Embodiment

• Robot: Fourier GR1 (robot_type: dream in meta/info.json)
• Action / state dim: 44 joints
• Camera: observation.images.ego_view

Joint layout (meta/modality.json):

Group	Indices
left_arm	0–6
left_hand	7–12
left_leg	13–18
neck	19–21
right_arm	22–28
right_hand	29–34
right_leg	35–40
waist	41–43

Directory layout (per subset)

{subset}_lerobot.data/
├── data/chunk-000/episode_XXXXXX.parquet
├── videos/chunk-000/observation.images.ego_view/episode_XXXXXX.mp4
└── meta/
    ├── info.json
    ├── episodes.jsonl
    ├── tasks.jsonl
    ├── modality.json
    └── stats.json

Each parquet row stores:

• observation.state — float32[44]
• action — float32[44]
• annotation.human.coarse_action — language-task index
• video frames referenced via LeRobot video paths

Note on actions. Videos and language labels come from Cosmos generation. The 44-D action / state columns follow the GR1 LeRobot schema. Inverse-dynamics (IDM) action labels (seonghyeonye/IDM_gr1) are intended as a follow-up upload under *_lerobot.data_idm/. Until that dump is published, treat action columns as schema placeholders unless you re-run IDM locally.

Language tasks

gr1_env

1. Use the right hand to pick up chip with green packaging to white bag
2. Use the right hand to pick up orange to metal platform
3. Use the right hand to pick up plastic pitcher and pour water onto green plant
4. Use the right hand to pick up the spoon from the bowl and serve the contents onto the plate
5. Use the right hand to knock over the green Pocky box

gr1_object

1. Use the left hand to pick up dark green cucumber from on circular gray mat to above beige bowl.
2. Use the right hand to pick up red bell pepper from center of tan countertop to brown bottom rack.
3. Use the left hand to pick up green pepper from tan table, below the bright blue plate to pale turquoise plate.
4. Use the right hand to pick up milk carton from pale turquoise plate in the center of the table to second level of wooden shelf.
5. Use the right hand to pick up green bok choy from top tier of two tier black and white shelf to brown paper bag.

gr1_behavior

1. Use the right hand to close lunch box
2. Use the right hand to pick up blue scoop and scoop powder from container
3. Use the right hand to pick up glass and bring it close to the camera as if drinking
4. Use knife to cut the object on the cutting board
5. Use the right hand to press calculator

How the data was produced

1. World-model videos — Cosmos-Predict2-14B Sample GR00T-Dreams-GR1, image + text conditioning, --disable_guardrail, 480p / 16 fps generation then resampled to 8 fps / 93 frames for Cosmos-Predict2 LeRobot packing.
2. Directory conversion — DreamGen Bench filenames → per-task folders.
3. LeRobot export — IDM_dump/raw_to_lerobot.py --cosmos_predict2.

Load with Hugging Face Hub

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="hk239/Vi2Act",
    repo_type="dataset",
)
# e.g. local_dir/gr1_env_lerobot.data/

Load as a LeRobot dataset

from gr00t.data.dataset import LeRobotSingleDataset

dataset = LeRobotSingleDataset(
    dataset_path="gr1_env_lerobot.data",  # after download
    embodiment_tag="gr1",
)

Citation

If you use this data, please cite DreamGen / GR00T-Dreams:

@article{jang2025dreamgen,
  title={DreamGen: Unlocking Generalization in Robot Learning through Video World Models},
  author={Jang, Joel and Ye, Seonghyeon and Lin, Zongyu and Xiang, Jiannan and Bjorck, Johan and Fang, Yu and Hu, Fengyuan and Huang, Spencer and Kundalia, Kaushil and Lin, Yen-Chen and others},
  journal={arXiv preprint arXiv:2505.12705v2},
  year={2025}
}

Acknowledgements

• nvidia/EVAL-175 — prompt / first-frame images
• nvidia/Cosmos-Predict2-14B-Sample-GR00T-Dreams-GR1 — video world model
• seonghyeonye/IDM_gr1 — inverse dynamics model (for action extraction)