Datasets:
metadata
pretty_name: AgiBot World
size_categories:
- 100K<n<2000K
task_categories:
- other
language:
- en
tags:
- real-world
- dual-arm
- Robotics manipulation
Agibot World Colosseo (Agibot) 🌍🤖
- Challenges with Current Benchmarks:
- Low data quality and limited sensing capabilities.
- Short-horizon tasks in controlled environments.
- First Large-Scale Robotic Learning Platform for Multi-purpose Robotic Manipulation.
- Includes data, models, benchmarks, and ecosystem.
- Aims to democratize real-robot data for the academic community.
- Aspires to trigger the “ImageNet moment” for Embodied AI in the near future.
Key Features 🔑
- One Million+ Demonstrations from 100 robots.
- 100+ real-world scenarios across 5 business sectors.
- Tasks involving:
- Fine-grained manipulation
- Long-horizon planning
- Dual-robot collaboration
- Cutting-Edge Hardware:
- Visuotactile Sensors
- 6-DoF hands
- Mobile dual-arm robots with full-body control
Research Potential 🌱
- Supports multimodal imitation learning and multi-agent collaboration, and beyond.
- State-of-the-art hardware for scalable robotic systems in production.
- Researchers and practitioners are invited to leverage this newly open-source platform to shape the future of Embodied AI.
Platform Release 📅
- Agibot Alpha: ~100,000 trajectories of real-robot data.
- Full platform suite coming by end of Q1 2025.
- Agibot-World Robot Manipulation Challenge
Get started
Download AgiBot Dataset
Installation
Dataset Structure
data
├── task_info
│ ├── task_374.json
│ ├── task_256.json
│ └── ...
├── 374[task_id]
│ ├── 64825[episode_id]
│ │ ├── depth
│ │ ├── parameters
│ │ ├── videos
│ │ ├── meta_info.json
│ │ └── aligned_joints.h5
│ ├── 67832
│ │ └── ...
│ └── ...
├── 321
│ └── ...
In the aligned_joints.h5 file, we organize the information in this format.
aligned_joints.h5
├── action
│ ├── joint
│ ├── effector
│ └── ...
├── state
│ ├── joint
│ ├── effector
│ └── ...