README.md

---
pretty_name: RoboInter-VQA
task_categories:
- robotics
- visual-question-answering
language:
- en
tags:
- Embodied-AI
- Robotic manipulation
- intermediate-representation
---

# RobotInter-VQA: Intermediate Representation Understanding & Generation VQA Dataset for Manipulation

**English** | [简体中文](./README_zh.md)

A Visual Question Answering dataset for robotic manipulation, developed as part of the [RoboInter](https://github.com/InternRobotics/RoboInter), covering **generation**, **understanding** (and **task planning**) of Intermediate Representations.  The dataset is built on top of the annotation of [RoboInter-Data](https://huggingface.co/datasets/InternRobotics/RoboInter-Data), raw robot datasets are from [DROID](https://droid-dataset.github.io/) and [RH20T](https://rh20t.github.io/).

## Dataset Structure

```
robotinter/
├── Generation/          # generation tasks (grounding, trajectory, contact, ...)
│   ├── image/           # Images (zip archives, extract before use)
│   │   ├── train/{droid,rh20t}/
│   │   └── val/
│   └── meta/            # VQA annotations in JSON
│       ├── train/{droid,rh20t}/{origin_format,llava_format,smart_resize_format}/
│       └── val/{origin_format,llava_format,smart_resize_format}/
├── Understanding/       # understanding tasks (multiple-choice/decision)
│   ├── image/
│   │   ├── train/{droid,rh20t}/
│   │   └── val/
│   └── meta/
│       ├── train/{droid,rh20t}/
│       └── val/
└── Task_planning/       # Task planning & primitive recognition
    ├── image/
    │   ├── train/manipvqa/
    │   └── val/{planning,choice,decide}/
    └── meta/
        ├── train/manipvqa/
        └── val/{planning,choice,decide}/
```

## Quick Start

1. **Extract images**: All images are stored as `.zip` files. Extract them in place:
   ```bash
   cd RoboInter-VQA/Task_planning/image/train/manipvqa
   cat task_planning.zip.* > task_planning_full.zip
   cd ../../../../
   find . -name "*.zip" -execdir unzip -o {} \;
   ```

2. **Load VQA data**: 
Refer to [RoboInterVLM](https://github.com/InternRobotics/RoboInter/blob/main/RoboInterVLM/RoboInterVLM-QwenVL/qwen-vl-finetune/qwenvl/data/__init__.py).


## Coordinate Formats (Generation only)

The Generation annotations are provided in **three coordinate formats**. The underlying data and images are identical; only the coordinate representation in the answers differs:

| Format | Description | Example |
|---|---|---|
| `origin_format` | Pixel coordinates in original image resolution (`h` x `w`) | `[[72, 102], [192, 179]]` |
| `llava_format` | Normalized coordinates in `[0, 1]` range as LLaVA-based Model | `[[0.22, 0.57], [0.60, 0.99]]` |
| `smart_resize_format` | Pixel coordinates in resized image resolution as Qwen-based Model (`new_h` x `new_w`) | `[[69, 95], [184, 167]]` |

## JSON Descriptions

### Generation (7 task types)

Each entry follows this schema:
```json
{
  "id": "unique_sample_id",
  "task": "task_type",
  "conversations": [{"from": "human", "value": "..."}, {"from": "gpt", "value": "..."}],
  "images": "relative/path/to/image.jpg",
  "gt": "ground_truth_value",
  "h": 180, "w": 320,
  "new_h": 168, "new_w": 308
}
```

| JSON File | Task | Description | Output Format |
|---|---|---|---|
| `*_traj_qa.json` | Trajectory Prediction | Given a task description and a **starting position**, predict 10 future trajectory waypoints for the gripper. | `{"future_traj": [[x1,y1], ...]}` |
| `*_traj_qa_wo_init_pos.json` | Trajectory Prediction (no init pos) | Same as above but **without** providing the starting position in the prompt. | `{"future_traj": [[x1,y1], ...]}` |
| `*_gripper_det_qa.json` | Gripper Detection | Detect the current bounding box of the robot gripper in the scene. | `{"gripper_det_bbox": [[x1,y1],[x2,y2]]}` |
| `*_contact_point_qa.json` | Contact Point Prediction | Predict the two contact points where the gripper fingers touch the manipulated object. | `{"contact_point": [[x1,y1],[x2,y2]]}` |
| `*_contact_box_qa.json` | Contact Box Prediction | Predict the bounding box of the gripper at the moment of contact with the object. | `{"contact_bbox": [[x1,y1],[x2,y2]]}` |
| `*_current_box_qa.json` | Current Object Box | Predict the current bounding box of the manipulated object. | `{"current_bbox": [[x1,y1],[x2,y2]]}` |
| `*_final_box_qa.json` | Final Object Box | Predict the final bounding box of the manipulated object (at the end of the action). | `{"final_bbox": [[x1,y1],[x2,y2]]}` |

### Understanding (6 task types)

Multiple-choice VQA tasks that evaluate visual understanding of intermediate representations. Each entry uses single-image or multi-choice concatenated images.

| JSON File | Task | Description | Answer |
|---|---|---|---|
| `contact_decide.json` | Contact Decision | Given a scene, determine whether the gripper has reached/contacted the object. | `Yes` / `No` |
| `grasppose_choice.json` | Grasp Pose Choice | Select the correct grasping pose from 4 candidate images (A/B/C/D), where orange fork-like patterns represent possible gripper poses. | `A`/`B`/`C`/`D` |
| `grounding_choice.json` | Grounding Choice | Select which image correctly depicts the bounding box (purple box) of the manipulated object from 4 candidates. | `A`/`B`/`C`/`D` |
| `traj_choice.json` | Trajectory Choice | Select the correct gripper trajectory from 4 candidate images with gradient-colored paths (green=start, red=end). | `A`/`B`/`C`/`D` |
| `trajlang_choice.json` | Trajectory-Language Choice | Given a trajectory visualization, select the correct task description from 4 language options. | `A`/`B`/`C`/`D` |
| `traj_direction_choice.json` | Trajectory Direction Choice | Given colored arrows around the gripper, select which color represents the actual movement direction. | `A`/`B`/`C`/`D` |

### Task Planning (16 task types)

Multi-image (video frame) or single-frame VQA tasks for high-level task planning, including 16 types (Scene understanding, Discriminative affordance negative task, Discriminative affordance positive task, Future multitask selection, Future prediction task, Future primitive selection task, Generative affordance task, Past description task, Past multitask selection, Past primitive selection, Planning remaining steps task, Planning task, Planning with context task, Success negative task, Success positive task, Temporal understanding). Each entry uses 8 sampled frames as input (Temporal understanding and Scene understanding use a combined image of four image). We provide 4 examples below:

| JSON File | Task | Description | Answer |
|---|---|---|---|
| `train/manipvqa/task_planning.json` | Next Step Planning (train) | Given 8 video frames and a goal, predict the next sub-task to perform. | Free-form text |
| `val/planning/task_planning.json` | Next Step Planning (val) | Same as training but on held-out validation data. | Free-form text |
| `val/choice/task_planning.json` | Primitive Selection (val) | Given 8 video frames, select which action primitive was just executed from 4 options. | `A`/`B`/`C`/`D` |
| `val/decide/task_planning.json` | Success Decision (val) | Given 8 video frames and a sub-task description, determine whether the sub-task was successfully completed. | `Yes` / `No` |

## Data Statistics

### Generation

| Source | traj_qa | gripper_det | contact_point | contact_box | current_box | final_box |
|---|---|---|---|---|---|---|
| DROID (train) | 31,282 | 84,777 | 78,004 | 78,004 | 149,671 | 145,996 |
| RH20T (train) | 33,803 | 120,747 | 115,266 | 115,266 | 225,055 | 224,944 |
| Val | 2,000 | 2,000 | 2,000 | 2,000 | 2,000 | 2,000 |

### Understanding

| Source | contact_decide | grasppose_choice | grounding_choice | traj_choice | trajlang_choice | traj_direction |
|---|---|---|---|---|---|---|
| RH20T (train) | 15,060 | 9,835 | 8,158 | 3,610 | 3,610 | 3,729 |
| DROID (train) | 18,184 | - | 57,572 | 8,245 | 8,245 | 6,500 |
| Val | 15,514 | 2,702 | 6,108 | 787 | 1,474 | 266 |

### Task Planning

| Split | Entries |
|---|---|
| Train (manipvqa) | 928,819 |
| Val - planning | 10,806 |
| Val - choice | 15,059 |
| Val - decide | 10,629 |

## Citation

If you find RoboInter useful in your research, please consider citing:

```bibtex
@article{li2026robointer,
  title={RoboInter: A Holistic Intermediate Representation Suite Towards Robotic Manipulation},
  author={Li, Hao and Wang, Ziqin and Ding, Zi-han and Yang, Shuai and Chen, Yilun and Tian, Yang and Hu, Xiaolin and Wang, Tai and Lin, Dahua and Zhao, Feng and Liu, Si and Pang, Jiangmiao},
  journal={arXiv preprint arXiv:2602.09973},
  year={2025}
}
```

## License

Please refer to the original dataset licenses for [RoboInter](https://github.com/InternRobotics/RoboInter), [DROID](https://droid-dataset.github.io/) and [RH20T](https://rh20t.github.io/).