puheliang/fmc3_gr2_grab_bottle_into_box_lerobot_ds

---
task_categories:
- robotics
- imitation-learning
language:
- en
license: mit
size_categories:
- 100K<n<1M
configs:
- config_name: default
data_files:
- split: train
path: data/chunk-000/file-000.parquet
---

# fmc3_gr2_grab_bottle_into_box_lerobot_ds

LeRobot-format manipulation dataset for Fourier GR2.

## Task

Pick up the bottle from the grid cell and place it into the box.

## Summary

- Robot: `fourier_gr2`
- Format: `LeRobot v3`
- FPS: `30`
- Total episodes: `332`
- Total frames: `145641`
- Split: `train`

## Sensors

This dataset contains six synchronized video streams:

- `observation.images.camera_top`
- `observation.images.camera_top_depth`
- `observation.images.camera_left_wrist`
- `observation.images.camera_left_wrist_depth`
- `observation.images.camera_right_wrist`
- `observation.images.camera_right_wrist_depth`

The top camera and both wrist cameras provide RGB + depth observations.

## Action / State

- `action`: 35D
- `observation.state`: 45D

## Source

Converted from an internal Dora-record multi-camera dataset collected on Fourier GR2.

- Source dataset name: `fmc3_gr2_grab_bottle_into_box_dora_ds`
- Converted with: `convert_parquet_to_lerobot_fourier_multicam.py`

## Conversion Notes

- RGB and depth streams are aligned by nearest timestamp resampling at 30 FPS.
- Depth streams are stored as 3-channel pseudo-RGB videos for compatibility with the current LeRobot video writer.
- Video files are chunked. Multiple episodes may share the same mp4 file and are indexed through `meta/episodes/chunk-000/file-000.parquet`.

## Data Quality Note

A total of 333 raw episodes were scanned during conversion.

- Kept: `332`
- Skipped: `1`
- Skipped episode: `episode_000000138`
- Skip reason: no common timestamp overlap across all required streams

## Directory Structure

```text
fmc3_gr2_grab_bottle_into_box_lerobot_ds/
├── meta/
├── data/
└── videos/
```

## Usage

This dataset is intended for LeRobot-compatible training and evaluation pipelines.