Notebookcheck Logo

AgiBot releases free humanoid robot training dataset

Free AgiBot World Alpha robotic learning dataset accelerates AI humanoid development. (Image source: AgiBot)
Free AgiBot World Alpha robotic learning dataset accelerates AI humanoid development. (Image source: AgiBot)
The AgiBot World Alpha dataset includes trained actions such as folding clothes. Over 100 scenarios, ranging from home to work environments, were trained with humans demonstrating correct outcomes. AI humanoid developers can use the dataset for free.

AgiBot has released the AgiBot World Alpha dataset for AI humanoid robot training and development. The dataset is provided under the Creative Commons CC BY-NC-SA 4.0 license, allowing for free, non-commercial use.

Some humanoid AI robots such as the 1X Neo have achieved the ability to learn from humans in actual situations to complete tasks such as cleaning rooms. The data used to power these humanoids is often proprietary and unavailable, which creates a digital divide between well-funded companies and the general public. AgiBot seeks to democratize the research and development of humanoids by releasing its high-quality robotic training dataset.

The dataset was created with humans guiding robots to the correct outcome in over 100 scenarios, such as folding clothes or pouring coffee. Challenging tasks that take more time and require long-horizon planning across five environments from home to office work were included. This opens the door to developing intelligent humanoids that can assist users with everyday tasks, including building a custom PC, grocery shopping, or packing shipping orders.

The AgiBot Genie-1 general purpose robot powered by the Nvidia Jetson AGX Orin embedded AI computer was used to collect the training data. The robot's arms have seven degrees-of-freedom and can lift 11 lbs. (5 kg) each. Multiple cameras on the robot monitor its surroundings as it rolls from task-to-task with a four-hour runtime.

AI robot developers wanting to download the AgiBot World Alpha dataset will need lots of free space since it occupies several hundred gigabytes. The dataset also depends on the LeRobot library. A fast external flash drive, like this 8 TB Samsung SSD on Amazon, or a slower external hard drive, like this 24 TB Western Digital HDD on Amazon, can be used to quickly expand storage space for AI developers using laptops.

AgiBot World Alpha includes high-quality humanoid training data across five environments and over 100 scenarios, with the majority of scenarios incorporating long-horizon planning. (Image source: AgiBot)
AgiBot World Alpha includes high-quality humanoid training data across five environments and over 100 scenarios, with the majority of scenarios incorporating long-horizon planning. (Image source: AgiBot)
The AgiBot Genie-1 general purpose robot powered by the Nvidia Jetson AGX Orin embedded AI computer was used to create the humanoid training dataset. (Image source: AgiBot)
The AgiBot Genie-1 general purpose robot powered by the Nvidia Jetson AGX Orin embedded AI computer was used to create the humanoid training dataset. (Image source: AgiBot)

Leading robotics startup AgiBot releases by far the largest humanoid manipulation dataset, paving the way for general-purpose robots in everyday life

News provided by AgiBot 

Dec 30, 2024, 07:49 ET

SHANGHAI, Dec. 30, 2024 /PRNewswire/ -- AgiBot, a leading robotics startup today releases by far the largest humanoid manipulation dataset: AgiBot World.

AgiBot World, the first large-scale robotic learning dataset specifically designed to advance multi-purpose robotic policies. This comprehensive ecosystem includes not only the dataset but also foundational models, standardized benchmarks, and a collaborative framework aimed at democratizing access to high-quality robotic data. It provides an unprecedented opportunity for both academia and industry to collaborate, paving the way for the "ImageNet Moment" in Embodied AI—a transformative leap toward universal, adaptable robotic intelligence.

Most existing robot learning benchmarks face significant limitations when addressing real-world challenges. These issues primarily stem from low-quality data and restricted sensing capabilities, resulting in benchmarks that are often constrained to short-horizon tasks within controlled environments. Such limitations hinder progress toward generalizable and robust robotic systems capable of operating effectively in unstructured, dynamic real-world settings.

With more than 1M trajectories from 100 robots, AgiBot World offers unprecedented diversity and complexity. Spanning more than 100 real-world scenarios across five target domains, it tackles fine-grained manipulation, tool usage, and multi-robot collaboration. These scenarios are meticulously designed to reflect the nuanced demands of real-world robotic applications.

Featuring cutting-edge multimodal hardware, AgiBot World provides array-based visual tactile sensors, durable 6-DoF hands, and mobile dual-arm robots with whole-body control. We hope these features will open new frontiers for research in areas such as multimodal imitation learning, multi-agent collaboration, adaptive manipulation, and beyond.

AgiBot World aspires to transform large-scale robot learning and advance scalable robotic systems for production. This open-source platform invites researchers and practitioners to collaboratively shape the future of Embodied AI.

For more information, please visit:

AgiBot World:https://agibot-world.com/ 

About AgiBot:

Established in February 2023, AgiBot is an innovative humanoid robot company. With the mission of "Create unlimited productivity via intelligent machines", AgiBot is dedicated to in-depth AI + robot fusion, aiming to create world-class leading embodied intelligent robot products and application ecosystems. Up to now, AgiBot has completed multiple rounds of financing, including the angel round and A1 - A4 rounds.

SOURCE AgiBot

static version load dynamic
Loading Comments
Comment on this article
Please share our article, every link counts!
Mail Logo
David Chien, 2024-12-30 (Update: 2025-01- 3)