(Yicai) Aug. 14 -- As orders for humanoid robots surge, putting them to work in China’s factories still faces technical challenges, according to industry insiders, who point out that a balance has yet to be struck between improving operational efficiency at plants and enhancing robot capabilities.

Although orders have soared this year, humanoid robots have not necessarily begun working in factories successfully, a source in the embodied intelligence sector told Yicai. Even if they did, it does not mean developers can obtain valuable data to train and upgrade embodied brain models, the person added.

Shanghai Jiao Tong University’s Prof. Gao Feng used the action of bolt screwing as an example, saying that contact-style tasks such as this are still very hard for industrial androids, and many technical problems need to be overcome. Mistakes, quality issues, and accidents are likely to occur if these issues are not resolved, Gao pointed out.

The developers that have landed large orders so far this year are Agibot, Unitree Robotics, and UBTech Robotics. Agibot secured a CNY78 million (USD10.9 million) order from a unit of China Mobile at the end of June, Unitree a CNY46.1 million order from the same firm, and UBTech a CNY90.5 million project from a car exporter last month.

UBTech expects to deliver about 500 industrial humanoid robots this year, mainly for handling, sorting, and quality inspection, Vice President Jiao Jichao said at the World Robot Conference that wrapped up two days ago. The Shenzhen-based firm counts auto, 3C (computers, communications, and consumer), and semiconductor companies among its clients, he added.

While the large language model and vision-language multimodal model aspects of embodied intelligence technology are developing rapidly and vision-language-action models are being explored, the tasks that almost all androids carry out today are the simplest, according to Liu Yunhui, a professor at the Chinese University of Hong Kong's engineering faculty. For example, precision grasping is often difficult to achieve, Liu noted.

Still, these challenges have not stopped this year from becoming the first for robot mass production, according to the insiders. But robot makers and their clients need to find the right balance between improving plant efficiency and upping robot performance, they added.

Clients want robots that can help production lines complete more complex work without having to remodel them, Xie Tiandi, marketing director of RoboSense Technology, told Yicai, adding that the embodied intelligence sector and its corporate customers must seek an equilibrium between enhancing operational efficiency at factories and the cost of introducing robots.

For example, adjusting the degrees of freedom of a robot's dexterous hand to an appropriate level that allows it to complete basic tasks in factories is sufficient for now, Xie pointed out.

More and more manufacturing clients no longer pursue robot technical capability for its own sake, but focus on cost-effectiveness, said Ma Yang, general manager of Tashi Technology. For instance, they no longer demand excessively high degrees of freedom for a robot’s dexterous hand, because they know that higher degrees demand more computing power, he said.

Editors: Tang Shihua, Martin Kadiev