Joocjun/GR1-Tabletop-NextState-1000x24

---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- GR00T
- GR1
- robocasa
- simulation
- manipulation
size_categories:
- 10K<n<100K
---

# GR1 Tabletop Merged LeRobot Datasets

Merged and subsampled versions of the GR1 tabletop manipulation datasets from the [NVIDIA PhysicalAI-Robotics-GR00T-X-Embodiment-Sim](https://huggingface.co/datasets/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](https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-Sim)**

To download the original dataset:
```bash
# 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

```python
# 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.