WorldAgents-c/MultiDojo2

MultiDojo²

English · 简体中文 · Project website · GitHub

Overview

2,000+ simulation demonstrations across 20+ multi-robot manipulation tasks. The collection covers homogeneous and heterogeneous teams of 2, 3, 4, and 8 arms, including handovers, relays, collaborative transport, ordered stacking, and parallel manipulation. Data is released progressively on this repository.

Every episode includes synchronized first-person views for every arm, third-person views, robot states and actions, and camera calibration. Multi-station tasks also include station views. RGB streams retain their recorded 640×480 resolution at 30 Hz.

Download and load

pip install -U huggingface_hub
hf download WorldAgents-c/MultiDojo2 --repo-type dataset --local-dir multidojo-data
cd multidojo-data
pip install -r requirements.txt

The currently downloadable files total approximately 98.1 GiB. Loading requires Python and the packages in requirements.txt; a simulator or GPU is unnecessary.

from loader import TrainingDataset

dataset = TrainingDataset("train.jsonl", max_open=2)
sample = dataset.step(episode=0, frame=0, action_horizon=8)
print(sample["instruction"])
print(sample["images"].keys())
print(sample["robots"].keys())
dataset.close()

For a partial download, include loader.py, requirements.txt, and every file in each selected episode directory, then filter train.jsonl to those episodes. Paths are relative to the index file. Directory numbers preserve the collection's original numbering and are non-contiguous; tasks.json maps every task to its episode directories and index rows.

Files

• train.jsonl and tasks.json: accepted episodes, task instructions, view names, and task-to-directory mapping.
• episodes/NNNN/render/lmdb/data.mdb: synchronized RGB, state, action, object-pose, and timing arrays.
• episodes/NNNN/render/meta_info.pkl: robot joint metadata, camera descriptions, and LMDB key mapping.
• episodes/NNNN/render/camera-calibration.npz: camera world poses and calibration parameters.
• episodes/NNNN/trajectory/trajectory.npz: native numeric trajectories, object poses, and robot link transforms.

Actions are absolute native joint-position commands, with radians or metres specified by each robot's joint metadata. Observation/state at t precede action t; next observations/state correspond to t+1. Variable joint dimensions, passive masks, and action chunks are supported. See FORMAT.md for details.

Tasks

ID	Task	Demonstrations
00	Banana handover into a pan	18
01	Banana and apple to plates	18
02	Collaborative carton transport	17
03	Two-stage cup relay	16
04	Place and stack blocks	18
05	Franka–Piper long-block handover	21
06	Three-arm block relay	11
07	Apples to baskets	11
08	Four-arm block relay	20
09	Ordered four-color stacking	10
10	Cookies to boxes	10
11	Household objects to boxes	18
12	Eight-arm blocks to boxes	14
13	Three-arm block relay	17
14	Four-arm block relay	14
15	Eight-arm blocks to boxes	11
16	Three-arm banana relay	14
17	Three-arm white-cup relay	11
18	Four-arm banana relay	20
19	Four-arm white-cup relay	12

Attribution

These are simulation recordings. Simulator executables and source robot/object assets are not bundled. See THIRD_PARTY_ATTRIBUTIONS.md and RIGHTS_STATUS.md.

中文说明

MultiDojo² 数据集包含 2,000+ 条训练轨迹、20+ 类多机器人协作任务，覆盖同构与异构的双臂、三臂、四臂和八臂组合。训练数据在本仓库持续开放。

每条轨迹同步保存每个机械臂的第一视角、第三视角、状态与动作，以及相机标定；多工位任务还包含工位视角。RGB 保持原始 640×480、30 Hz。当前可下载文件约 98.1 GiB。

使用 train.jsonl 加载，tasks.json 提供每类任务对应的目录和索引行。目录编号沿用原始编号，因此不连续。安装 requirements.txt 后可使用 loader.py 读取。状态动作的单位、同步关系和仿真建模说明见 FORMAT.md。