AgiBotWorldChallenge-2026

# AgiBot World Challenge 2026 - Datasets

Dear participants,

We are excited to announce that the datasets for AgiBot World Challenge 2026 have been updated.

# Track 1: Reasoning2Action

The Reasoning to Action track evaluates models' capabilities in reasoning and action prediction, comprising both online and onsite phases. This track encompasses 10 progressively challenging tasks, ranging from basic to complex, including dual-arm collaboration, long-horizon operations, and high-precision manipulations such as logistics sorting, office organization, retail operations, and daily services.

Based on G2 robot, AgiBot World open datasets, and Genie Sim 3.0, this track focuses on bridging the Sim2Real gap and achieving robust generalization from open-vocabulary understanding to physical interaction.

## Competition Tasks

| No. | Task Name |
| --- | --- |
| 1 | clean_the_desktop |
| 2 | hold_pot |
| 3 | open_door |
| 4 | place_block_into_box |
| 5 | pour_workpiece |
| 6 | scoop_popcorn |
| 7 | sorting_packages |
| 8 | sorting_packages_continuous |
| 9 | stock_and_straighten_shelf |
| 10 | take_wrong_item_shelf |

## Dataset Structure
The dataset repository contains three main directories at the top level (to lower the barrier to entry, we also provided a version of the same dataset without depth):
```
Reasoning2Action-Sim/
├── clean_the_desktop_part_1/
│ ├── data.tar.gz.000
│ ├── meta.tar.gz.000
│ ├── videos.tar.gz.000
│ ├── ...
│ └── videos.tar.gz.013
├── clean_the_desktop_part_2/
│ ├── data.tar.gz.000
│ ├── meta.tar.gz.000
│ ├── videos.tar.gz.000
│ ├── ...
│ └── videos.tar.gz.011
├── ... (Other task folders have the same structure as above.)
└── >>> dataset_without_depth/ <<<
├── clean_the_desktop_part_1/
│ ├── data.tar.gz.000
│ ├── meta.tar.gz.000
│ ├── videos.tar.gz.000
├── clean_the_desktop_part_2/
│ ├── data.tar.gz.000
│ ├── meta.tar.gz.000
│ ├── videos.tar.gz.000
└── ... (Other task folders have the same structure as above.)
```

> **dataset_without_depth: version without depth data**

The directory structure after decompression should be as follows:
```
{task name}/
├── meta/
│ ├── episodes.jsonl
│ ├── episodes_stats.jsonl
│ ├── info.json
│ └── tasks.jsonl
├── data/
│ └── chunk-000/
│ ├── episode_000000.parquet
│ ├── ...
│ └── episode_000400.parquet
└── videos/
└── chunk-000/
├── observation.images.hand_left/
├── observation.images.hand_right/
├── observation.images.top_head/
├── observation.images.hand_right_depth/
├── observation.images.hand_left_depth/
└── observation.images.head_depth/
```

# 2026年AgiBot世界挑战赛（AgiBot World Challenge 2026）数据集

亲爱的参赛者：
我们很高兴地宣布，2026年AgiBot世界挑战赛的数据集已完成更新。

## 赛道1：推理到行动（Reasoning2Action）
推理到行动赛道旨在评估模型的推理与行动预测能力，赛事分为线上与线下两个阶段。该赛道包含10项难度逐级递增的任务，覆盖从基础到复杂的各类场景，涵盖双臂协作、长时序操作以及高精度操控等类型，具体涉及物流分拣、办公整理、零售作业与日常服务等领域。
本赛道基于G2机器人（G2 robot）、AgiBot世界公开数据集以及Genie Sim 3.0平台打造，核心目标在于弥合仿真到现实（Sim2Real）的差距，并实现从开放词汇（open-vocabulary）理解到实体交互的鲁棒泛化能力。

### 竞赛任务
| 编号 | 任务名称 |
| --- | --- |
| 1 | 清理桌面（clean_the_desktop） |
| 2 | 持握锅具（hold_pot） |
| 3 | 开门（open_door） |
| 4 | 将积木放入盒中（place_block_into_box） |
| 5 | 倾倒工件（pour_workpiece） |
| 6 | 铲取爆米花（scoop_popcorn） |
| 7 | 包裹分拣（sorting_packages） |
| 8 | 连续包裹分拣（sorting_packages_continuous） |
| 9 | 补货并整理货架（stock_and_straighten_shelf） |
| 10 | 从货架取走错误商品（take_wrong_item_shelf） |

### 数据集结构
数据集仓库顶层包含三个主要目录（为降低入门门槛，我们还提供了不含深度数据的同版本数据集）：

Reasoning2Action-Sim/
├── clean_the_desktop_part_1/
│ ├── data.tar.gz.000
│ ├── meta.tar.gz.000
│ ├── videos.tar.gz.000
│ ├── ...
│ └── videos.tar.gz.013
├── clean_the_desktop_part_2/
│ ├── data.tar.gz.000
│ ├── meta.tar.gz.000
│ ├── videos.tar.gz.000
│ ├── ...
│ └── videos.tar.gz.011
├── ...（其余任务文件夹的结构与上述一致。）
└── >>> dataset_without_depth/ <<<
├── clean_the_desktop_part_1/
│ ├── data.tar.gz.000
│ ├── meta.tar.gz.000
│ ├── videos.tar.gz.000
├── clean_the_desktop_part_2/
│ ├── data.tar.gz.000
│ ├── meta.tar.gz.000
│ ├── videos.tar.gz.000
└── ...（其余任务文件夹的结构与上述一致。）

> **dataset_without_depth：无深度数据的数据集版本**

解压后的目录结构如下所示：

{task name}/
├── meta/
│ ├── episodes.jsonl
│ ├── episodes_stats.jsonl
│ ├── info.json
│ └── tasks.jsonl
├── data/
│ └── chunk-000/
│ ├── episode_000000.parquet
│ ├── ...
│ └── episode_000400.parquet
└── videos/
└── chunk-000/
├── observation.images.hand_left/
├── observation.images.hand_right/
├── observation.images.top_head/
├── observation.images.hand_right_depth/
├── observation.images.hand_left_depth/
└── observation.images.head_depth/