AgiBotWorldChallenge-2025

# AgiBot World Challenge 2025 - Datasets

<img src="poster.jpg" alt="Image Alt Text" width="100%" style="display: block; margin-left: auto; margin-right: auto;" />

Dear participants,

We are excited to announce that the datasets for both tracks of **AgiBot World Challenge 2025** have been updated.

---

## Changelog
Previous versions remain available in the branch version name.
- v1.1.0(11-08-2025): Added 50 new trajectories per Manipulation task and synced all datasets to [ModelScope](https://modelscope.cn/datasets/AgiBotWorld/agibot_world_challenge_2025/files).
- v1.0.1 (27-06-2025): Fixed an Issue with the annotations in the simulation data of the "open drawer and store items" task.
- v1.0.0 (25-06-2025): Initial version. Released datasets for both tracks of AgiBot World Challenge 2025.

## Track 1：Manipulation

We have specifically collected data for 10 distinct tasks for this competition, with hundreds of trajectories per task. Utilizing an advanced data collection approach - Adversarial Data Collection (ADC), we've incorporated dynamic disturbances to significantly enhance the information density and diversity of each trajectory. This approach not only reduces post-training data requirements and model training costs but also effectively strengthens the model's generalization capabilities and robustness.

The **simulation** datasets for 10 tasks are provided in the Manipulation-SimData folder. To facilitate local simulation evaluation for participants, we have open-sourced a set of **simulation assets** (path: https://huggingface.co/datasets/agibot-world/GenieSimAssets).

Each **real-robot dataset** has one or two corresponding files, and here is the mapping between task names and their respective dataset IDs:

| Task Name | Real Robot Dataset ID |
|----------------------------------|------------------------------|
| Heat the food in the microwave | 881 |
| Open drawer and store items | 949, 1019 |
| Pack in the supermarket | 1352, 1418 |
| Stamp the seal | 1458 |
| Pack washing detergent from conveyor | 1645 |
| Clear the countertop waste | 1957 |
| Pickup items from the freezer | 1918, 2055 |
| Restock supermarket items | 1967 |
| Make a sandwich | 1969 |
| Clear table in the restaurant | 1968 |

---

## Track 2：World Model

We've designed a series of challenging tasks covering scenarios including kitchen environments, workbenches, dining tables, and bathroom settings, encompassing diverse robot\-object interactions \(e\.g\., collisions, grasping, placement, and dragging maneuvers\), to thoroughly evaluate models' generative capabilities\.

This track offers a comprehensive dataset consisting of training, validation, and testing sets:

### Dataset Structure

- **Training Set**
The training set includes over __30,000__ premium trajectories selected from __10__ representative tasks in the AgiBot World Dataset, providing ample material for model training\.

- **Validation Set**
The validation set contains __30__ carefully chosen samples to support model verification and optimization\.

- **Testing Set**
The testing set includes 30 no-public samples, covering both seen and unseen scenarios in the training set, mixed with expert demonstrations and imperfect trajectories, aiming to assess models' generalization and robustness comprehensively\.

```
DATASET_ROOT/
├── train/
│ ├── 367-648961-000/
│ │ ├── head_color.mp4
│ │ ├── head_extrinsic_params_aligned.json
│ │ ├── head_intrinsic_params.json
│ │ └── proprio_stats.h5
│ ├── 367-648961-001/
│ │ ├── head_color.mp4
│ │ ├── head_extrinsic_params_aligned.json
│ │ ├── head_intrinsic_params.json
│ │ └── proprio_stats.h5
│ ├── {task_id}-{episode_id}-{step_id}/
│ │ ├── head_color.mp4
│ │ ├── head_extrinsic_params_aligned.json
│ │ ├── head_intrinsic_params.json
│ │ └── proprio_stats.h5
│ └── ...
├── val/
│ ├── 367-649524-000/
│ │ ├── head_color.mp4
│ │ ├── head_extrinsic_params_aligned.json
│ │ ├── head_intrinsic_params.json
│ │ └── proprio_stats.h5
│ └── ...
└── test/
├── {task_id}-{episode_id}-{step_id}/
│ ├── frame.png
│ ├── head_color.mp4 (NOT disclosed to participants)
│ ├── head_extrinsic_params_aligned.json
│ ├── head_intrinsic_params.json
│ └── proprio_stats.h5
└── ...
```

### Provided Data Includes

- EEF poses
- Joint angles
- Camera intrinsics/extrinsics
- ......

→ Enabling participants to fully utilize physical and visual information.

---

We look forward to seeing your innovative solutions in the challenge!

# 2025年AgiBot世界挑战赛（AgiBot World Challenge 2025）数据集

<img src="poster.jpg" alt="Image Alt Text" width="100%" style="display: block; margin-left: auto; margin-right: auto;" />

尊敬的参赛选手：
我们很高兴宣布，**2025年AgiBot世界挑战赛**双赛道的数据集均已完成更新。

---

## 更新日志
旧版本仍可在对应分支版本中获取。
- v1.1.0（2025-08-11）：为每个操控任务新增50条新轨迹，并将所有数据集同步至[ModelScope](https://modelscope.cn/datasets/AgiBotWorld/agibot_world_challenge_2025/files)平台。
- v1.0.1（2025-06-27）：修复了“打开抽屉并存放物品”任务的仿真数据标注问题。
- v1.0.0（2025-06-25）：初始版本。发布2025年AgiBot世界挑战赛双赛道的官方数据集。

---

## 赛道1：操控（Manipulation）
本次竞赛针对10项差异化任务专门采集了数据集，每项任务包含数百条轨迹。我们采用先进的数据采集方法——对抗性数据采集（Adversarial Data Collection, ADC），通过引入动态扰动显著提升了每条轨迹的信息密度与多样性。该方法不仅降低了模型训练所需的数据量与训练成本，还有效增强了模型的泛化能力与鲁棒性。

10项任务的**仿真（simulation）数据集**已存放于Manipulation-SimData文件夹中。为方便参赛选手进行本地仿真评估，我们开源了一套**仿真资源（simulation assets）**，访问地址为：https://huggingface.co/datasets/agibot-world/GenieSimAssets。

每项**实体机器人数据集（real-robot dataset）**对应1至2个相关文件，任务名称与对应数据集ID的映射关系如下表所示：

| 任务名称 | 实体机器人数据集ID |
|-----------------------------------|--------------------|
| 用微波炉加热食物 | 881 |
| 打开抽屉并存放物品 | 949、1019 |
| 超市打包作业 | 1352、1418 |
| 盖章作业 | 1458 |
| 从传送带上收纳洗涤剂 | 1645 |
| 清理台面杂物 | 1957 |
| 从冷冻柜取放物品 | 1918、2055 |
| 补货超市商品 | 1967 |
| 制作三明治 | 1969 |
| 清理餐厅餐桌 | 1968 |

---

## 赛道2：世界模型（World Model）
我们设计了一系列覆盖厨房环境、工作台、餐桌及浴室场景的挑战性任务，涵盖丰富的机器人-物体交互行为（例如碰撞、抓取、放置及拖拽操作），以全面评估模型的生成能力。

本赛道提供包含训练集、验证集与测试集的完整数据集：
### 数据集结构
- **训练集**：从AgiBot世界数据集的10项典型任务中精选了超过30000条优质轨迹，可为模型训练提供充足的数据支撑。
- **验证集**：包含30份精心挑选的样本，用于支持模型验证与优化。
- **测试集**：包含30份未公开的样本，覆盖训练集中出现过的及全新的场景，混合了专家演示轨迹与非完美轨迹，旨在全面评估模型的泛化能力与鲁棒性。

DATASET_ROOT/
├── train/
│ ├── 367-648961-000/
│ │ ├── head_color.mp4
│ │ ├── head_extrinsic_params_aligned.json
│ │ ├── head_intrinsic_params.json
│ │ └── proprio_stats.h5
│ ├── 367-648961-001/
│ │ ├── head_color.mp4
│ │ ├── head_extrinsic_params_aligned.json
│ │ ├── head_intrinsic_params.json
│ │ └── proprio_stats.h5
│ ├── {task_id}-{episode_id}-{step_id}/
│ │ ├── head_color.mp4
│ │ ├── head_extrinsic_params_aligned.json
│ │ ├── head_intrinsic_params.json
│ │ └── proprio_stats.h5
│ └── ...
├── val/
│ ├── 367-649524-000/
│ │ ├── head_color.mp4
│ │ ├── head_extrinsic_params_aligned.json
│ │ ├── head_intrinsic_params.json
│ │ └── proprio_stats.h5
│ └── ...
└── test/
├── {task_id}-{episode_id}-{step_id}/
│ ├── frame.png
│ ├── head_color.mp4（不向参赛选手公开）
│ ├── head_extrinsic_params_aligned.json
│ ├── head_intrinsic_params.json
│ └── proprio_stats.h5
└── ...

### 提供的数据包含
- 末端执行器位姿（End-Effector Poses, EEF poses）
- 关节角度（Joint angles）
- 相机内参/外参
- ......
→ 助力参赛选手充分利用物理与视觉信息。

---

我们期待在本次挑战赛中见证各位选手的创新解决方案！