BlueRun Ventures Headlines | AgiBot Open-Sources Million-Scale Real-Robot Dataset! The ImageNet Moment for Embodied AI Has Arrived!
The world's first million-real-robot dataset built on full-coverage real-world scenarios, all-capability hardware platforms, and end-to-end quality control
As a key partner in Shanghai's "Model Shanghai" AI implementation plan, AgiBot has joined forces with the Shanghai Artificial Intelligence Laboratory, the National-Local Joint Engineering Research Center for Humanoid Robotics, and Shanghai Corpus to officially launch AgiBot World today — the world's first million-scale real-robot dataset built on comprehensive real-world scenarios, universal hardware platforms, and end-to-end quality control. This landmark open-source project marks the arrival of the "ImageNet moment" for embodied intelligence.
BlueRun Ventures is an early investor in AgiBot and has participated in three consecutive funding rounds.
This is AgiBot's third open-source project this year, and it fully delivers on the commitment made at the AgiBot 818 launch event. We will upload the data in batches to HuggingFace, GitHub, and the agibot-world.com project page according to plan, accelerating humanity's march toward artificial general intelligence.
Project Links
HuggingFace: https://huggingface.co/agibot-world
GitHub: https://github.com/OpenDriveLab/agibot-world
Project page: https://agibot-world.com/
AgiBot World is the world's first million-scale real-robot dataset built on comprehensive real-world scenarios, universal hardware platforms, and end-to-end quality control. Compared to Google's Open X-Embodiment dataset, AgiBot World features 10x more long-horizon data, 100x broader scenario coverage, and quality elevated from laboratory-grade to industrial-grade standards! In this world of embodied data, robots no longer perform simple tabletop tasks — they enter every aspect of human daily life.
The scenarios covered in the AgiBot World dataset are remarkably diverse, spanning from basic operations like grasping, placing, pushing, and pulling to complex actions such as stirring, folding, and ironing — encompassing nearly all scenarios required for human daily life. In one video, for instance, a robot installs a RAM module into a computer chassis, requiring millimeter-level precision control where the slightest error could damage the equipment. Its nerve-fiber-sensitive tactile sensor at the end effector enables precise alignment. Another video attempts to teach a robot how to use a dishwasher — in this data sample, spoons, chopsticks, bowls, and plates are stacked in layers in the kitchen sink, and the robot methodically organizes the cluttered utensils into the corresponding slots of the dishwasher, a lengthy and intricate operation. The AgiBot World dataset includes over 80 diverse skill videos from daily life, enabling robots to master "eighteen martial arts."
The AgiBot World dataset was born in AgiBot's self-built large-scale data collection factory and application testing base, with a total floor area exceeding 4,000 square meters and containing over 3,000 real-world items. On one hand, it provides facilities for large-scale robot data training; on the other, it faithfully replicates five core scenarios — home, dining, industrial, retail, and office — comprehensively covering typical application needs for robots in both production and daily life.
The home scenario reproduces authentic residential layouts, including bedrooms, living rooms, kitchens, and bathrooms, enabling robots to perform household cleaning, item organization, and kitchen tasks. The industrial scenario simulates sorting and logistics automation, replicating industrial warehouses and production lines with sorting systems, packaging equipment, and conveyor belts, enabling material sorting, packaging, and logistics handling. Through highly faithful scenario reconstruction and task design, AgiBot World builds the necessary conditions for achieving embodied intelligence in robot R&D and testing. There are over 100 such real-world scenarios in the AgiBot World million-scale real-robot dataset, with home accounting for 40%, dining 20%, industrial 20%, retail 10%, and office 10%. Eighty percent of tasks are long-horizon, with task durations concentrated between 60-150 seconds and comprising multiple atomic skills — long-horizon data exceeds that of DROID and Open X-Embodiment by more than 10x. The 3,000+ items basically cover these five scenarios and continue to expand.
Scenario and task distribution
Item categories by scenario
Dataset duration distribution
That robots can be so agilely intelligent and competent across multiple tasks and skills also owes to AgiBot's iterative upgrades to the robot hardware itself — we equipped the robot with eight cameras in a surround configuration, enabling real-time 360-degree perception of dynamic environmental changes; a 6-DOF dexterous hand ensures precise and flexible movement, while the end effector adds a six-axis force/torque sensor and high-precision tactile sensor capable of perceiving minute force variations, achieving "measured delicacy"; the robot's full body has 32 active degrees of freedom, enabling flexible handling of various complex tasks.
Meanwhile, dataset quality is also key to whether robots can learn quickly. All data in AgiBot World must undergo a rigorous process and verification chain. From the project's inception, we invited input from academia, industry, and consumers alike, continuously iterating on process design and planning. For data collection quality, both collectors and collection quality are ensured by a comprehensive management system and professional management team throughout the process. For the collected data itself, strict screening is applied on both edge and cloud sides, automatically filtering out non-compliant data, followed by frame-by-frame review by professional auditors to ensure every action meets task standards. Finally, the data undergoes secondary verification through algorithms, building comprehensive defenses for data quality on all fronts.
High-quality datasets are particularly important for the current development of embodied intelligence technology. Existing open-source datasets more or less suffer from issues such as lack of standardized collection processes, outdated robot configurations, and inconsistent data quality and formats, which can even cause harm during robot policy learning.
Google's OXE dataset: tasks lack real-life scenarios, diverse robot forms, inconsistent data quality and formats
AgiBot's open-sourcing of AgiBot World — the world's first million-scale real-robot dataset built on comprehensive real-world scenarios, universal hardware platforms, and end-to-end quality control — represents a major breakthrough and milestone for embodied intelligence technology, and an important accelerator for advancing embodied AGI. By bringing together top-tier resources and technical expertise, we will jointly drive a new paradigm in embodied intelligence development, accelerate humanity's march toward artificial general intelligence, and establish China's leadership in this frontier field on the global stage.
Going forward, AgiBot will successively open-source tens of millions of simulation data points to support more generalizable and universal large model training; release an embodied foundation model supporting fine-tuning to empower industries across the board; and release a complete toolchain achieving a closed loop of collection, training, and evaluation.
Recapping Shenzhen's hottest AI hardware salon last week | Booming Hub recap
BlueRun Ventures and portfolio companies win multiple awards including Most Influential Investment Institution of the Year
Sinorythm successfully completes Series A+ funding round | BlueRun Ventures
