README.md

# MiGA: Multi-gripper dataset for Gripper-aware Vision-Language-Action models

[![License](https://img.shields.io/badge/License-Apache_2.0-yellow.svg)](https://opensource.org/licenses/Apache-2.0)
[![Hugging Face](https://img.shields.io/badge/Hugging_Face-MiGA--Dataset-000000.svg)](https://huggingface.co/datasets/GVLA/MiGA-Dataset)
[![arXiv](https://img.shields.io/badge/arXiv-2608.24603-red.svg?style=flat-square)](http://arxiv.org/html/2608.24603v1)
[![ECCV 2026](https://img.shields.io/badge/ECCV-2026-purple.svg)](#)

Companion dataset for **GVLA (Gripper-aware Vision-Language-Action Models)**.

---

## Overview of MiGA
MiGA is a large-scale **multi-gripper-aware** dataset featuring diverse gripper types on complex tasks, explicitly capturing how the same task requires different strategies depending on the gripper morphology.

- **Format:** [LeRobot](https://github.com/huggingface/lerobot) dataset schema, stored as Parquet
- **Domains:** real-world + simulation
- **Modalities:** third-person RGB (`image`), wrist-camera RGB (`wrist_image`), robot `state` (pose + gripper opening, 8-dim), `actions` (delta pose + gripper command, 7-dim), per-episode/task indices
- **License:** Apache 2.0

### Composition of MiGA

MiGA is released as a large-scale multi-gripper-aware dataset featuring diverse gripper types on complex tasks, explicitly capturing how the same task requires different strategies depending on the gripper morphology. MiGA is built on five different gripper types (parallel-jaw, vacuum/suction, 3-finger, dexterous hand, soft gripper). Subsets are grouped below by domain (real-world vs. simulation).

**Real-world subsets**

| Gripper type | Gripper config | Robot platform | Episodes | Format | HF dataset repo |
|---|---|---|---|---|---|
| Parallel-jaw | Franka panda hand | Franka Panda | 1877 | Parquet (LeRobot) | [GVLA/Franka_panda_parallel_hand_real](https://huggingface.co/datasets/GVLA/Franka_panda_parallel_hand_real) |
| Vacuum | Cobot Pump | Franka + Cobot | 671 | Parquet (LeRobot) | [GVLA/Franka_cobot_vacuum_real](https://huggingface.co/datasets/GVLA/Franka_cobot_vacuum_real) |
| Dexterous hand | Inspire dexterous hand | Franka + Inspire Hand | 899 | Parquet (LeRobot) | [GVLA/Franka_inspire_hand_real](https://huggingface.co/datasets/GVLA/Franka_inspire_hand_real) |
| Parallel-jaw | Robotiq 85 parallel-jaw | UR5 | 106 | Parquet (LeRobot) | [GVLA/UR5_robotiq_85_parallel_real](https://huggingface.co/datasets/GVLA/UR5_robotiq_85_parallel_real) |
| 3 finger | Robotiq 3-f | UR5 | - | Parquet (LeRobot) | [GVLA/UR5_robotiq_3f_3finger_real](https://huggingface.co/datasets/GVLA/UR5_robotiq_3f_3finger_real) |
| Soft Gripper | Soft-fin | X-arm7 | - | Parquet (LeRobot) | coming soon |

**Simulation subsets**

| Gripper type | Gripper config | Robot platform | Episodes | Format | HF dataset repo |
|---|---|---|---|---|---|
| Parallel-jaw | Franka panda hand | Franka Panda | 4254 | Parquet (LeRobot) | [GVLA/Franka_panda_parallel_hand_sim](https://huggingface.co/datasets/GVLA/Franka_panda_parallel_hand_sim) |
| Parallel-jaw | Robotiq 85 parallel-jaw  | Franka Panda | 4060 | Parquet (LeRobot) | [GVLA/Franka_robotiq_85_parallel_sim](https://huggingface.co/datasets/GVLA/Franka_robotiq_85_parallel_sim) |
| Vacuum | Cobot | Franka Panda | 4824 | Parquet (LeRobot) | [GVLA/Franka_cobot_vacuum_sim](https://huggingface.co/datasets/GVLA/Franka_cobot_vacuum_sim) |
| Vacuum | Short-cup vacuum | Franka Panda | 6025 | Parquet (LeRobot) | [GVLA/Franka_short_cup_vacuum_sim](https://huggingface.co/datasets/GVLA/Franka_short_cup_vacuum_sim) |
| Vacuum | Short-cup vacuum | UR10 | 2949 | Parquet (LeRobot) | [GVLA/UR10_short_cup_vacuum_sim](https://huggingface.co/datasets/GVLA/UR10_short_cup_vacuum_sim) |

<!-- > Episode counts are the number of distinct `episode_index` values in each subset (i.e. `max(episode_index) + 1`), not the frame-level row counts shown in the Hugging Face viewer summary. Confirmed so far: `Franka_panda_parallel_hand_sim` (239 episodes, `episode_index` 0–238) and `Franka_panda_parallel_hand_real` (90 episodes, `episode_index` 0–89). The remaining five are marked `[FILL IN]` — open each subset's Data Studio viewer and check the max value shown under the `episode_index` column header (e.g. "episode_index int64  0  89"), or check `meta/info.json`'s `total_episodes` field if using the LeRobot loader locally.
>
> **Note:** `GVLA/Franka_panda_hand_parallel_sim` also appears in the org with the same row count (63.2K) as `GVLA/Franka_panda_parallel_hand_sim` — these look like they may be the same data under two different repo names. Worth double-checking whether one is a duplicate/leftover before publishing, so the card doesn't list it twice.
 -->
<!-- ### Task categories

Each gripper subset is organized by `task_index`, with per-episode `episode_index` grouping and frame-level `frame_index`/`timestamp`. The parallel-jaw simulation subset spans multiple grasping tasks (task indices 0–4 observed); the real-world parallel-jaw subset currently reflects a smaller task set in its viewer sample. [FILL IN: give the full task list / task names per subset — e.g. pick-and-place, articulated-object manipulation, tool use.]

### Diversity of objects & scenes

[FILL IN: 1–2 sentences — object categories, scene types (tabletop/kitchen/etc.), number of unique objects across all subsets.] -->

---

## Data description

Each subset follows the standard **LeRobot** dataset layout (Parquet data files + `meta/info.json`), auto-converted by Hugging Face into a browsable Parquet dataset:

```
GVLA/<gripper_config_repo>/
|-- data/
|   `-- train-*.parquet        # episodes concatenated, indexed by episode_index/frame_index
|-- videos/                    # camera MP4s (if published separately from the parquet)
|-- meta/
|   `-- info.json              # robot type, fps, feature schema
```

### Per-frame fields (confirmed from the live dataset viewer)

| Field | Type | Description |
|---|---|---|
| `image` | image (224×224) | Third-person RGB frame |
| `wrist_image` | image (224×224) | Wrist-camera RGB frame |
| `state` | float32[8] | End-effector pose (xyz + rotation) + gripper opening (2 values) |
| `actions` | float32[7] | Delta end-effector action + gripper command |
| `gripper_id` | int32 | Identifies which gripper config generated this episode (e.g. `0` = parallel-jaw sim, `2` = parallel-jaw real) |
| `timestamp` | float32 | Time within episode (s) |
| `frame_index` | int64 | Frame index within episode |
| `episode_index` | int64 | Episode index within subset |
| `index` | int64 | Global row index |
| `task_index` | int64 | Task identifier |

## Example of data usage

```python
from datasets import load_dataset

# Load one gripper-config subset directly
ds = load_dataset("GVLA/Franka_panda_parallel_hand_real", split="train")
print(ds[0]["gripper_id"], ds[0]["state"], ds[0]["actions"])

# Or with the LeRobot dataset loader (recommended for training)
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("GVLA/Franka_panda_parallel_hand_real")
```

To combine subsets for cross-gripper training, concatenate on the shared `state`/`actions`/`gripper_id` schema and use `gripper_id` to condition or stratify by embodiment.

---

## Version update

### Version 1.0
Initial release across seven gripper/domain subsets: Franka parallel-jaw (sim + real), Franka+Cobot vacuum (real), Franka+Inspire hand (real), UR5+Robotiq 85 parallel-jaw (real), UR5+Robotiq 3-finger (real), and UR10+short-cup vacuum (sim).

---

## Citation

If you find MiGA useful in your research, please cite:

```bibtex
@inproceedings{zhang2026gvla,
  title     = {Gripper-aware Vision Language Action Models},
  author    = {Hanyi Zhang, Zihong Luo, Tianyu Li, Khang Nguyen, Basu Hela, Shreyas Kumar, Ngoc Duy Tran, Feng Dai, Charith Munasinghe, Jorge Peña Queralta, Giovanni Toffetti, Khoa Vo, Ngan Le, Ravi Prakash, Quan Vuong, Tung D. Ta, Long Hu, Anh Nguyen, and Baoru Huang},
  booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
  year      = {2026},
  eprint    = {2608.24603},
  archivePrefix = {arXiv}
}
```

Paper: [arxiv.org/html/2608.24603v1](http://arxiv.org/html/2608.24603v1)

## Reference

- Project page: [https://airvlab.github.io/G-VLA/]

<!-- ## Contact / Discussions
For questions or issues, please open a Discussion on any of the dataset repos above, or contact [FILL IN: email]. -->