AgiBot, or Zhiyuan Robotics, a Chinese robotics firm aimed at building and innovating humanoid robotics today announced it’s releasing what it says is the largest robotic learning dataset for advancing artificial intelligence foundation models for human-like activities.

Founded in 2023 by Peng Zhihui, the Shanghai-based startup competes with Tesla Inc. and Boston Dynamics Inc. in the race to build and deliver humanoid robots that can perform basic tasks similar to how humans can. For example, things such as folding clothing, organizing items, cleaning and doing laundry.

“Most existing robot learning benchmarks struggle to address real-world challenges caused by low-quality data and limited sensing capabilities, typically limited to short-horizon tasks within controlled environments,” the company said about the dataset.

To address this problem AgiBot released a large-scale, open-source dataset called AgiBot World, which contains more than 1 million diverse training sets from 100 robots covering what the company says is an extremely diverse set of activities. The training set covers more than 100 real-world scenarios across five target domains, including fine-grained manipulation, tool usage and multi-robot collaboration.

The data contained within includes 40% home-related tasks, 20% from restaurants, 20% industrial, 10% office work and 10% supermarket-related tasks. The company explained it trained robots in numerous fine-grained and long-term planning-oriented tasks including picking up, moving, opening, closing, stacking, turning and more in order to emulate human movements. Robots trained on the dataset can also readily collaborate with other human workers in group tasks, not just other robots.

Compared to other open datasets, such as Google LLC’s Open X-Embodiment, AgiBot said AgiBot World has 10 times more long-range navigational data and covers 100 times more scenarios for humanoid robots. It also used more real-world training in industrial-grade environments rather than data produced purely in a laboratory environment.

To train humanoid robots, the source, accuracy and precision of data is paramount, especially regarding task-level data. Robotic AI can be trained using real-world data for handling numerous activities based on a number of sensors such as cameras, lidar, touch and more. This allows AI models to plan out motions based on what the robot can see and understand in its environment – however, the training set needs to be extremely comprehensive to cover many different objects, issues and cases.

For example, AgiBot’s training set includes data for allowing bots to put together a PC motherboard, move dishes in a sink and even work together with other humanoid bots to move a table across a room.

The company has also built its own “data collection factory” where it gathers real-world data on various industry, domestic and everyday tasks using practical training. The company said that it can even provide enterprise-quality custom data on request for training and fine-tuning robotics AI models.

Since its launch, AgiBot has introduced humanoid and wheeled robots designed for industrial and domestic applications.

AgiBot World is available on HuggingFace and GitHub for robotics AI developers and academic researchers to use and shape their own foundation AI models for humanoid robots.

Image: AgiBot