chaoqi-liu/robocasa_mt5_PH100_TH500

---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- robocasa
- robocasa365
- multitask
configs:
- config_name: default
data_files: data/*/*.parquet
pretty_name: robocasa_mt5_PH100_TH500
---

# robocasa_mt5_PH100_TH500

This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).

## Dataset Description

This is a custom **RoboCasa365 v1.0.0** dataset exported in LeRobot format.
It was built from the `mt5` 5-task group using human demos only.

`mt5` is the 5-task group used for this dataset.
It expands to the five RoboCasa leaf tasks listed below.

### Name Breakdown

- `robocasa_mt5_PH100_TH500`
- `mt5` = the 5-task group used for this dataset
- `PH100` = up to `100` demos per task from the `pretraining-human` pool (`pretrain-human`)
- `TH500` = up to `500` demos per task from `target-human`

### The 5 Tasks

- `CloseToasterOvenDoor`
- `OpenDrawer`
- `PickPlaceDrawerToCounter`
- `TurnOnElectricKettle`
- `SlideDishwasherRack`

### Per-task Episode Counts

- `CloseToasterOvenDoor`: `PH=100`, `TH=500`, total `600` episodes
- `OpenDrawer`: `PH=100`, `TH=500`, total `600` episodes
- `PickPlaceDrawerToCounter`: `PH=100`, `TH=500`, total `600` episodes
- `TurnOnElectricKettle`: `PH=100`, `TH=500`, total `600` episodes
- `SlideDishwasherRack`: `PH=100`, `TH=500`, total `600` episodes

That yields `3000` episodes total.

### What `total_tasks = 20` Means

The `mt5` group contains **5 RoboCasa env tasks**, but `meta/info.json` reports
`total_tasks = 20` because `meta/tasks.parquet` stores the full vocabulary of
task-related strings used by the dataset, including:

- canonical task names such as `CloseToasterOvenDoor`
- human instruction strings such as `Close the toaster oven door.`
- object-specific instruction variants for `PickPlaceDrawerToCounter`
- side/direction variants for tasks like `OpenDrawer` and `SlideDishwasherRack`

The frame data use integer indices, and those indices resolve through `meta/tasks.parquet`
for fields like `annotation.human.task_name`, `annotation.human.task_description`, and
`task_index`.

## Dataset Structure

[meta/info.json](https://huggingface.co/datasets/chaoqi-liu/robocasa_mt5_PH100_TH500/blob/main/meta/info.json):

```json
{
"codebase_version": "v3.0",
"robot_type": "PandaOmron",
"total_episodes": 3000,
"total_frames": 639196,
"total_tasks": 20,
"fps": 20,
"splits": {
"train": "0:3000"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4"
}
```

- [meta/stats.json](https://huggingface.co/datasets/chaoqi-liu/robocasa_mt5_PH100_TH500/blob/main/meta/stats.json)
- [meta/tasks.parquet](https://huggingface.co/datasets/chaoqi-liu/robocasa_mt5_PH100_TH500/blob/main/meta/tasks.parquet)
- [meta/episodes/chunk-000/file-000.parquet](https://huggingface.co/datasets/chaoqi-liu/robocasa_mt5_PH100_TH500/blob/main/meta/episodes/chunk-000/file-000.parquet)

## Source

- Underlying benchmark family: `RoboCasa365 v1.0.0`
- Dataset repo: `chaoqi-liu/robocasa_mt5_PH100_TH500`
- Pool caps used to build this merge: `{'pretrain-human': 100, 'target-human': 500}`

## Notes

- This dataset includes `pretrain-human` and `target-human` demos only.
- `pretrain-mimicgen` was not included in this merge.
- Total size: `639196` frames at `20` FPS.