Joocjun/GR1-Tabletop-NextState-100x24

GR1 Tabletop Merged LeRobot Datasets

Merged and subsampled versions of the GR1 tabletop manipulation datasets from the NVIDIA PhysicalAI-Robotics-GR00T-X-Embodiment-Sim collection, formatted in LeRobot v2.0 format.

Dataset Variants

Variant	Demos/Task	Tasks	Total Episodes	Total Frames	Approx Size
1000x24/	1000	24 folders, 186 unique tasks	24,000	6,020,058	~40 GB
300x24/	300	24 folders, 186 unique tasks	7,200	1,803,236	~12 GB
100x24/	100	24 folders, 185 unique tasks	2,400	602,846	~4 GB

Source Dataset

The original per-task datasets were downloaded from: nvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-Sim

To download the original dataset:

# Install huggingface_hub
pip install huggingface_hub

# Download specific task datasets
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="nvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-Sim",
    repo_type="dataset",
    allow_patterns="gr1_unified.PnP*",  # or specific task patterns
    local_dir="./original_data"
)

The 24 source task folders used (each with 1000 demos):

• 6 PnP tasks: Bottle→Cabinet, Can→Drawer, Cup→Drawer, Milk→Microwave, Potato→Microwave, Wine→Cabinet
• 18 PosttrainPnPNovel tasks: Various object transfers between surfaces (Cuttingboard→{Basket,Cardboardbox,Pan,Pot,Tieredbasket}, Placemat→{Basket,Bowl,Plate,Tieredshelf}, Plate→{Bowl,Cardboardbox,Pan,Plate}, Tray→{Cardboardbox,Plate,Pot,Tieredbasket,Tieredshelf})

Processing

Each merged dataset was created by:

1. Subsampling: Randomly sampling N episodes (with seed=42) from each of the 24 task folders
2. Merging: Combining all sampled episodes into a single dataset with:
   • Re-indexed episode IDs (0 to total-1)
   • Globally re-indexed task IDs across all source folders
   • Updated parquet files with corrected episode_index, task_index, and index columns
   • Consolidated meta/ files (info.json, episodes.jsonl, tasks.jsonl, modality.json)

Note on task counts

The 100x24 variant has 185 unique tasks instead of 186. This is because some task folders contain multiple task variants (e.g., the PosttrainPnPNovelFromTrayToTieredbasket folder has 10 different object variants like eggplant, tomato, etc.). The "eggplant" variant only appears in 58 out of 1000 source episodes, so when randomly sampling only 100 episodes, none of the eggplant episodes were selected. This is expected behavior with random subsampling.

Dataset Structure (LeRobot v2.0)

<variant>/
├── data/
│   └── chunk-000/
│       ├── episode_000000.parquet
│       ├── episode_000001.parquet
│       └── ...
├── videos/
│   └── chunk-000/
│       └── observation.images.ego_view/
│           ├── episode_000000.mp4
│           ├── episode_000001.mp4
│           └── ...
└── meta/
    ├── info.json
    ├── episodes.jsonl
    ├── tasks.jsonl
    ├── modality.json
    ├── stats.json
    └── relative_stats.json

Features

Feature	Type	Shape
observation.images.ego_view	video (h264)	[256, 256, 3] @ 20fps
observation.state	float64	[44]
action	float64	[44]
timestamp	float64	[1]
next.reward	float64	[1]
next.done	bool	[1]
task_index	int64	[1]
episode_index	int64	[1]
annotation.human.coarse_action	int64	[1]

Robot: GR1ArmsAndWaistFourierHands (44-DOF: arms, hands, legs, neck, waist)

Usage

# Download a specific variant
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Joocjun/GR1-Tabletop-Merged-LeRobot",
    repo_type="dataset",
    allow_patterns="100x24/**",  # or "300x24/**" or "1000x24/**"
    local_dir="./gr1_merged"
)

License

Apache 2.0 (following the original NVIDIA dataset license)

Modification: Next-State Hand Actions

Hand action channels (left_hand indices 7-12, right_hand indices 29-34) have been replaced with next-state values:

• action[t][7:13] = observation.state[t+1][7:13]
• action[t][29:35] = observation.state[t+1][29:35]
• Last timestep repeats current state.
• stats.json removed so training auto-regenerates normalization stats.