China News Service, Hangzhou, October 17 (Xi Jinyan) Embodied intelligence is regarded as the next trillion-level industry after computers, smartphones, and new energy vehicles. Now in Zhejiang, many manufacturing companies have taken the lead in exploring the implementation of humanoid robot production lines, trying to let them undertake loading, transfer and other operations in real work stations. However, from pilot to large-scale "work in factories", multiple thresholds such as operating rhythm, stability, and mass production maturity still need to be crossed. How far are humanoid robots from being truly integrated into the production system? Zhejiang's manufacturing frontline is exploring answers.
Walking into Zhaofeng's electromechanical workshop in Xiaoshan, Hangzhou, 15 Leju humanoid robots are carrying out parts "loading" operations in an orderly manner. They move at a steady rhythm, and independently complete the entire process of identification, grabbing, transfer, and feeding in a small production line environment with metal frames and multiple workstations.
On the front line of the workshop, multiple robots are loading materials. Photo provided by interviewee
"Although they are still a little rusty now, they have made a lot of progress. They often fell down when they first came here. After 15,000 training iterations, they have achieved today's operating results." Recalling the scene of the introduction of robots for debugging and training in the factory a year ago, Kong Aixiang, chairman of Zhejiang Zhaofeng Mechanical and Electrical Co., Ltd., still remembers it fresh.
Talking about the application of embodied intelligence in factories, he is full of confidence in the prospects: "At present, the stability and accuracy of robots have basically reached the standards. In the future, we will improve the rhythm and other details to match the real scene. We hope to explore a set of industrial application solutions that are truly implementable and can be promoted throughout the industry."
Zhaofeng Electromechanical has been deeply involved in the field of automotive parts for more than 30 years and is a leading domestic automotive wheel bearing enterprise. On the basis of traditional automation transformation, the company took the lead in trying to embed large quantities of humanoid robots into real production stations, and jointly released the "Embodied Intelligence + Auto Parts Loading Solution" with Leju Robot.
Unlike fixed robotic arms, humanoid robots have a human-like torso and flexible arms, and can adapt to existing factory layouts without requiring large-scale transformation of production lines. This is also a core advantage that manufacturing companies value. However, the complex industrial site environment, random placement of materials, narrow space, and fluctuating working conditions are all realistic tests for the implementation of humanoid robots.
In the past, the embodied intelligence industry was mostly focused on single-point pilots and technical verification, with few normalized and replicable implementation cases. In the Zhaofeng electromechanical production line, 15 humanoid robots can complete more than 10 action sequences in a narrow and complex dynamic environment with interlaced wire mesh frames, multi-stations, and loading tables, stably grabbing small parts with an inner diameter of 80 to 120 mm. The time to complete a loading cycle is stably controlled within 40 seconds, the loading success rate is stable at 98%, and it has been running continuously without failure on site for more than 2 months.
Ke Zhendong, executive vice president of Leju Robot, said that 15 humanoid robots operate simultaneously on the Mega production line, forming large-scale parallel data collection capabilities. Calculated based on the current data production capacity, more than 40,000 items are collected in a single month, and continue to accumulate as the production line runs. These data all come from real industrial scenarios and can be used for model iteration to make robots more and more "smart".
It is worth mentioning that the two parties have donated the "automotive industry parts loading real machine operation reflow data set" to the Open Atom Open Source Foundation OpenLET community to enrich real machine data resources in industrial scenarios, lower the industry-wide research and development threshold, and help more scientific research teams and enterprises carry out embodied intelligent model training.
The person in charge of the Open Atom Open Source Foundation believes that the continuous accumulation of real operation data will also improve the generalization capabilities of robots. Based on the continuous accumulation of data and continuous optimization of models, combined with the skill development tool chain, multi-machine scheduling platform and universal dexterity, standard skills in the same scenario can be quickly adapted to different parts and different workstations. The modular design does not require large-scale customization and development, and can be quickly copied to other workstations in the same production line and other scenarios in the same industry.
Moving from a single scenario to multiple scenarios requires joint efforts from the industrial ecosystem. Ren Guangjie, assistant vice president of Leju Robots, said that in order to truly cover a large number of manufacturing scenarios, a large number of ecological teams in the upstream and downstream of the industry chain still need to jointly participate to jointly complete the capability puzzle in all aspects of implementation. The next step will be to unite more industry chain partners, continue to open source real machine data, development tool chains and other resources, improve the OpenLET embodied intelligent open source data set community construction, provide support for developers and innovative entrepreneurial teams, promote "verify one, deploy a batch, drive a group", and promote the implementation of more high-value industrial scenarios.
On the one hand, the real production line continues to accumulate data and polish the model, and on the other hand, the industrial scene has strict requirements for stability and economy. This experiment, which took place in a Zhejiang workshop, provides a real sample for the industrialization of humanoid robots. From pilot work to large-scale work in factories, humanoid robots have taken a critical step, but the long-distance race to industrialization has just begun. (over)
