---
license: apache-2.0
task_categories:
- robotics
tags:
- robotics
- lerobot
- so101
- isaac-lab
- code-as-policies
- multitask
- basecap
- 10fps
size_categories:
- 1M<n<10M
configs:
- config_name: default
  data_files: data/*/*.parquet
---

# IsaacLab SO101 11-Task baseCaP Dataset (3300 Episodes, 10 FPS)

LeRobot v3.0 dataset derived from
`CoRL2026-CSI/Isaaclab-so101_11task_baseCaP_3300epi` by keeping every third
episode-local source frame (`frame_index % 3 == 0`) and rebuilding each
episode's `frame_index`, `timestamp`, and global `index` for 10 FPS playback.

## Dataset Summary

| Field | Value |
|---|---:|
| Episodes | 3,300 |
| Frames | 1,175,352 |
| FPS | 10 |
| Source FPS | 30 |
| Source frames | 3,522,774 |
| Robot | `so101_follower` |
| Format | LeRobot v3.0 |
| Source dataset | `CoRL2026-CSI/Isaaclab-so101_11task_baseCaP_3300epi` |
| Source revision | `dba8cf9041a4ae3420987e53d9184f97a4d1cf86` |

## Tasks

This dataset contains 11 IsaacLab SO101 task families, 300 episodes each:

| Index | Task family | Episodes |
|---:|---|---:|
| 0 | Pick and place cube | 300 |
| 1 | Sort cubes by color | 300 |
| 2 | Stack two cubes | 300 |
| 3 | Push button | 300 |
| 4 | Turn lever on | 300 |
| 5 | Turn lever off | 300 |
| 6 | Extract cube | 300 |
| 7 | Open box | 300 |
| 8 | Close box | 300 |
| 9 | Push cube | 300 |
| 10 | Pull cube | 300 |

## Cameras

| Key | Shape | Codec | Description |
|---|---:|---|---|
| `observation.images.top` | 480 x 640 x 3 | h264 | Top-view RGB camera |
| `observation.images.left_wrist` | 480 x 640 x 3 | h264 | Left wrist RGB camera |

## Features

The dataset keeps the source schema, including:

- robot state/action fields in degree and URDF-radian forms
- end-effector pose and gripper state
- per-frame `skill.natural_language`, `skill.type`, and `skill.progress`
- per-skill goal joint, EE pose, and gripper fields
- per-frame `subtask.natural_language`, `subtask.object_name`, and `subtask.target_position`
- SCRAPE sidecar metadata in `meta/scrape_episode_metadata.jsonl`
- multi-task campaign metadata in `meta/scrape_multitask_manifest.json`

## Verification

The local derivative passed the SCRAPE dataset audit:

- `FAIL=0`, `WARN=0`, `PASS=37`
- 3,300 episodes and 1,175,352 data rows
- both camera video packet counts match `meta/info.json` total frames
- all per-episode video segment lengths match `meta/episodes`
- per-episode frame indices are contiguous from 0
- timestamps start at 0 and advance by 0.1 seconds
- sampled non-video columns match source rows selected by `frame_index % 3 == 0`

## Quick Start

```python
from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("CoRL2026-CSI/Isaaclab-so101_11task_baseCaP_3300epi_10fps")
sample = dataset[0]
print(sample.keys())
```