---
pretty_name: ManiSoft
license: mit
language:
- en
size_categories:
- 1K<n<10K
task_categories:
- robotics
tags:
- robotics
- soft-robotics
- manipulation
- imitation-learning
- vision-language-action
- embodied-ai
- simulation
---

# ManiSoft

ManiSoft is a soft-robot manipulation dataset and benchmark for vision-language-action learning. It contains expert demonstrations for four manipulation tasks:

- `COLL`: Collection
- `ALN`: Alignment
- `ARR`: Arrangement
- `STK`: Stacking

This upload directory currently provides:

- `assets.tar`: simulator assets required for replay and training
- `clean/`: task data packaged as `.tar` shards for efficient download and upload
- `data_extract.sh`: a utility script for recursively extracting all dataset shards

## Task Layout in This Repository

The files hosted in the dataset repository are organized as tar shards rather than already-extracted case folders.

```text
.
├── assets.tar
├── clean
│   ├── ALN
│   │   ├── train_bottle_0_9.tar
│   │   ├── train_bottle_10_19.tar
│   │   ├── eval_bottle_0_9.tar
│   │   └── ...
│   ├── ARR
│   │   ├── eval_bottle_0_9.tar
│   │   └── ...
│   ├── COLL
│   │   ├── train_pencup_0_9.tar
│   │   ├── eval_boxdrink_0_9.tar
│   │   └── ...
│   └── STK
│       ├── train_default_0_9.tar
│       ├── eval_default_0_9.tar
│       └── ...
└── data_extract.sh
```

For `ALN`, `ARR`, and `COLL`, shard names follow:

```text
<split>_<object_category>_<start_case_id>_<end_case_id>.tar
```

For `STK`, shard names follow:

```text
<split>_default_<start_case_id>_<end_case_id>.tar
```

## Extracted Dataset Format

After extraction, each shard restores the original directory structure. A typical case directory looks like this:

```text
clean/
└── ALN/
    ├── train/
    │   └── bottle/
    │       └── 0/
    │           ├── environment.yaml
    │           ├── instructions.txt
    │           ├── trajectory.pkl
    │           └── visual/
    └── eval/
        └── bottle/
            └── 0/
                ├── environment.yaml
                ├── instructions.txt
                ├── trajectory.pkl
                └── visual/
```

Each case is typically organized by:

```text
<setting>/<task>/<split>/<object_category>/<case_id>/
```

Common files inside one case:

- `instructions.txt`: language instructions for the manipulation case
- `environment.yaml`: scene and task configuration
- `trajectory.pkl`: expert trajectory stored as a time-indexed dictionary
- `visual/`: visualization assets such as rendered frames or videos

## Quick Download Example

If you use the Hugging Face CLI, you can download the dataset to a local directory like this:

```bash
hf download JobsWei/ManiSoft --local-dir ./ManiSoft --repo-type dataset
```

If you only need the benchmark data without simulator assets:

```bash
hf download JobsWei/ManiSoft --local-dir ./ManiSoft --repo-type dataset --exclude "assets.tar"
```

If you only need evaluation shards:

```bash
hf download JobsWei/ManiSoft --local-dir ./ManiSoft --repo-type dataset --include "**/eval/**"
```

## `data_extract.sh` Usage

The repository includes `data_extract.sh` for recursively finding and extracting all `.tar` files under a root directory with parallel workers.

### Command

```bash
bash data_extract.sh <tar_root_dir> <max_processes> <delete_tar_file>
```

### Arguments

- `tar_root_dir`: root directory to recursively search for `.tar` files
- `max_processes`: number of parallel extraction processes, must be a positive integer
- `delete_tar_file`: whether to delete each `.tar` after successful extraction
  - `0`: keep tar files
  - `1`: delete tar files

### Typical Examples

Extract all dataset shards under the downloaded directory and keep the original tar files:

```bash
bash data_extract.sh ./ManiSoft 8 0
```

Extract all dataset shards and delete each tar file after successful extraction:

```bash
bash data_extract.sh ./ManiSoft 8 1
```

Extract only the `clean` subset:

```bash
bash data_extract.sh ./ManiSoft/clean 8 1
```

### What the Script Does

- recursively finds all `.tar` files under `tar_root_dir`
- extracts them in parallel
- restores files into the original relative paths stored in each tar shard
- optionally removes the source tar files after successful extraction

## Recommended Workflow

```bash
hf download JobsWei/ManiSoft --local-dir ./ManiSoft --repo-type dataset --exclude "assets.tar"
cp /path/to/data_extract.sh ./ManiSoft/
cd ./ManiSoft
bash data_extract.sh ./clean 8 1
```

If you also need simulator assets:

```bash
tar -xvf assets.tar
```

## Notes

- The extraction script requires a Unix-like shell environment with `bash`, `find`, `tar`, and standard job control support.
- Different shards may expand into the same `train/` or `eval/` directory tree. This is expected.
- `trajectory.pkl` is the main expert trajectory file used for imitation learning and replay.