(Yicai) March 18 -- Fourier, a leading Chinese general-purpose robotics company, has open-sourced its android dataset Fourier ActionNet and the world's first full-process toolchain, including acquisition, training, and deployment algorithms.

Fourier aims to enhance artificial intelligence robot training and offer a comprehensive solution for global developers and research institutions, encouraging innovation and collaboration in the robotics community, the Shanghai-based firm said on WeChat yesterday.

The Fourier ActionNet dataset contains a variety of training tasks for the Fourier GRx humanoid robot series, including precise pick-and-place and dumping of commonly used tools, household items, food, and other objects, and generalized execution under different environmental conditions, the company added.

"High-quality datasets allow robots to learn operation methods and copy strategies in different scenarios, enhance generalization capabilities, and master various skills more comprehensively, significantly improving their learning ability and task execution efficiency," Zheng Yangyang, an AI robot researcher at Samoyed Cloud Technology Group Holding, told Yicai.

The dataset's quality impacts the efficiency of robot training, an insider from the embodied intelligence field said to Yicai, adding that regarding AI models' training requirements, the data usage rate of third parties may be less than 1 percent. "We often buy 1 million pieces of data, but after strict quality screening, only 10,000 or even fewer can be used."

The data in the Fourier ActionNet dataset is automatically annotated through a visual language model and manual secondary verification to ensure the accuracy of the data and training, a relevant person in charge of Fourier told Yicai.

"We will launch multiple datasets in the future, including for the robot's lower body movement," the person noted.

Most open-source datasets focus on hand operations and do not cover whole-body movement, interactive actions in complex environments, and other important tasks, Zheng pointed out. Datasets' generalization capabilities in complex environments still need improvement, Zheng said.

Editor: Martin Kadiev