Chinese humanoid robot market undergoes rapid price correction driven by automotive supply chain synergies while facing data bottlenecks and deployment limits
Chinese market for humanoid robots has begun to go through a rapid price correction process and, soon will become a low cost product instead of a high priced niche item. Now you can find 1,000,000 dollar humanoid robotics for 50,000 RMB (full truck of components) on a second and even on salvage markets. Some consumer and low cost robots can be bought for less than a smartphone; this shows that domestic manufacturing capacity is now overperforming current demand.
the Unitree G1 now starts from 85000 RMB, and the entry level R1 Air model is sold at the price of 29900 RMB. Another market player, Songyan Dynamics, is offering its Bumi humanoid for 9998 RMB, which is less than an iPhone. In the light of the current pricing crash, even renting humanoid hardware from the rental market is now cheaper. The rental price of humanoid robots fell from 10000 RMB a day to only several hundred RMB. Nevertheless, commercial customers find it difficult to pinpoint a clear, high value business case for deploying the new machines on a large scale.
The main factor enabling such a drastic price decline is China's powerful manufacturing sector, since humanoid robots and electric vehicles share much in terms of components and manufacturing processes. The physical joints, which are the most important physical component for robots, account for nearly 20,000 RMB of a 40,000 RMB robot, such as the Unitree G1. Chinese manufacturers are managing to localize between 75 90% of the key components of humanoid robots: actuators, controllers, gear reducers and servo systems.
This is the case for gear reducers too, one of the few key components that are currently dominated by Japanese manufacturers. Chinese engineers have successfully multiplied the operational lifetime of harmonic reducers by more than 3 times, reaching up to 10000 hours. Simultaneously, their price decreased to less than 33% of the original price. Chinese product developers have also achieved increased energy transmission efficiency compared with foreign products.
Despite the dramatically low price of hardware, its actual industrial application is extremely limited. According to research from Gartner, for each active robot, there are 60 units that are in the research exploration stage of deployment and only 1.64% of enterprises surveyed by Gartner have successfully applied the technology in their industrial processes. At the Zeekr 5G smart factory, the Zeekr robots with S1 industrial humanoids working on an assembly line can only be considered as the only exceptions for now.
Such low rates of deployment are questioning the practicality of human like robotics. Gao Ting, VP of Gartner, claims that
"human robots should innovate upon the human like shape rather than exactly imitate it"
suggesting that the body structure of robots should not be limited to resemble the human form. In example, Amazon tested Digit robot’s knees bend backwards so that it could be more steady in squatting positions to reach high shelves of warehouses, whereas the EVE robot by 1X uses a wheeled, self stabilized platform with significantly higher speed. Other non humanoid robots, such as quadruped robots or those based on wheel driven platforms, could potentially be more stable, faster and cheaper.
Besides hardware components, high quality operational data is the bottleneck in robotics development. Unlike large language models trained on Internet text and rich databases of words, robots are supposed to learn from physical interactions, which means robot experts need expensive teleoperation to acquire motion data, or use simulators for virtual training and then perform many experiments in the real environment. However, the reality is that physical world has countless variables that are difficult to be programmed using synthetic simulation data.
Gao Ting added that even if a robot can complete many simulated moves successfully in a virtual machine, its motion commands will often fail in reality due to variables like micro level frictions, changed material, or lighting conditions. This is called the "sim to real" gap. Moreover, direct training based on videos from online repositories still struggles with the so called "embodiment gap", i.e. The differences between the robot's physical embodiment and a human's physical capabilities make mapping the movements very inefficient.
To overcome these computational bottlenecks, the hybrid data strategy that places importance on physical interactions at the center, and combined with teleoperation and manually assisted data gathering, plus motion capture and synthetic simulation, should be prioritized by robotics developers. Although the ultra low prices of robots like the R1 Air reflect the immense power of Chinese manufacturing, the physical forms need to be accompanied with smart data in order for them to go from merely experimental products to intelligent working machines.
