yuuu94/W2-VLA-CoT

World-to-Wrist: Offline CoT Labels

This dataset contains frame-aligned offline chain-of-thought annotations used to train W²-VLA policies on LIBERO, RoboTwin, and four real-world manipulation tasks. Matching LeRobot action data is available in W2-VLA-Training-Data.

Dataset Structure

W2-VLA-CoT/
├── libero/
│   ├── libero_10_no_noops_1.0.0_lerobot/
│   ├── libero_goal_no_noops_1.0.0_lerobot/
│   ├── libero_object_no_noops_1.0.0_lerobot/
│   └── libero_spatial_no_noops_1.0.0_lerobot/
├── robotwin/
│   └── <50 task directories>/
├── real_world/
│   ├── place_bag/
│   ├── put_mango/
│   ├── table_clean/
│   └── plug_in_socket/
└── dataset_manifest.json

Split	Episodes	Frames
LIBERO	1,693	273,465
RoboTwin	2,500	549,787
Real world: place bag	80	24,000
Real world: put mango	100	31,000
Real world: table clean	100	89,469
Real world: plug in socket	100	75,679
Total	4,573	1,043,400

Annotation Format

Each episode_XXXXXX.npz file contains frame-aligned arrays. The policy is trained with cot_train_text, which has the following three-field format:

Subtask: ...
Reasoning: ...
Wrist: ...

The public schema contains:

schema_version
episode_index
num_frames
task
task_description
task_name
cot_train_text
cot_subtask
cot_reasoning
cot_wrist_focus

Task fields are present when available. The four CoT arrays have length num_frames.

Download

hf download yuuu94/W2-VLA-CoT \
  --repo-type dataset \
  --local-dir playground/Datasets/W2-VLA-CoT

Use libero/, robotwin/, or real_world/ as the label root for the matching training configuration.

License

The W2-VLA CoT annotations produced and curated by the W2-VLA authors are released under the Creative Commons Attribution 4.0 International License (CC BY 4.0). Users may use, copy, redistribute, adapt, clean, and modify the annotations, including using them as offline supervision for other VLA architectures. Appropriate attribution must be provided, and modified versions must indicate that changes were made.

This license applies to the W2-VLA annotation layer and does not replace the licenses governing any underlying benchmark data or assets.