New Gains in PRC Robotics Software & Hardware

Publication: China Brief Volume: 25 Issue: 19

By:

Age: 10 hours

A humanoid robot, developed by Shenzhen-based UBTech, operating in a BYD factory. (Source: CnEVPost)

Executive Summary:

  • A humanoid robotics ecosystem is developing in the People’s Republic of China (PRC). A vertically integrated supply chain is improving global competitiveness and market share in core component technologies.
  • In most segments, Western firms remain dominant. PRC firms note that they are constrained by foreign dependence on certain technologies, machine tools, and data quality.
  • Software, simulation, and multimodal AI remain particular chokepoints. But reliance on Nvidia’s physical AI stack and foreign cloud infrastructure is substantial.

Humanoid robotics made headlines throughout the People’s Republic of China (PRC) in September with a bold push to collect high-quality training data. Across key hubs in Beijing, Henan, and Jiangsu, the country is steadily cultivating firms and building supply chains that can rival global leaders (Shanghai Observer, September 30). National and provincial-level policies promote the localization of hardware production and seek to reduce vulnerabilities stemming from reliance on imports. Advances in sensor technologies, as well as in the component parts of robots’ “muscles” are notable. For now, Western companies dominate many parts of the value chain, including in machine tools. But Chinese firms are starting to erode these positions. In cases, such as software, the use of U.S. systems is allowing companies to advance rapidly.

Catching up with Leading Western Sensor Firms

Humanoid robots depend on sensors for basic functionality. These come in two types. Self-sensing sensors tell the robot what its own body is doing. World-sensing sensors, meanwhile, refer to devices that recognize objects and map environments. Robots rely on complex analysis of integrated data points from each of these sensors to perform basic activities.

In the PRC, national policy has supported the growth of humanoid robots. Central government directives have labeled “bionic sensors” (仿生传感) as a priority, backing the sector with funding and pilots to localize supply chains (International Center for Science & Technology Innovation [NCSTI], November 3, 2023). Well-known Chinese vendors like Hesai Technology (禾赛科技) and RoboSense Technology (速腾聚创) now lead the global automotive LiDAR market. Other Chinese firms also have made gains. But because they are less frequently addressed in policy discussions, they have flown largely under the radar (Yole Group, April  11).

Changchun Yuheng Optics (长春禹衡) is the leading domestic manufacturer of rotary encoders—devices that sit at a robot’s knee or elbow and ensures it knows exactly how far the joint has moved. Yuheng Optics is a national “little giant” (小巨人) enterprise, and currently comprises about 9 percent of domestic market share (Jiuyan Gongshe, December 22, 2024; Changchun Evening News, May 21). Nearly half of the domestic market remains dominated by foreign enterprises, though, with Japan’s Tamagawa Seiki and Germany’s Heidenhain the leaders. These firms also set international standards for encoders (Baijiahao, November 4, 2024).

A similar story exists for force/torque sensors. U.S.-based ATI Industrial Automation continues to lead the global market in the most advanced kind of torque sensor, the six-dimensional sensor. By capturing both force and torque in three directions, 6D sensors are especially useful for complex tasks like handling delicate objects. Chinese firms are starting to catch up. Leading domestic companies include Keli Sensing (柯力传感), Sunrise Instruments (SRI; 宇立仪器), Link-touch (蓝点触控), and Changzhou Kunwei Sensing Technology (Kunwei; 坤维科技). Kunwei is now able to produce sensors at 30 percent lower costs than ATI equivalents while maintaining quality for humanoid and surgical robotics (Xueqiu, May 11). Link-touch’s share of the PRC’s humanoid robot 6D torque sensor market surpassed 70 percent in 2024. And in the first half of 2025, it shipped 70,000 units, accounting for over 95 percent of total domestic shipments (Zhihu, July 21).

Western firms also have an edge in the development of tactile sensors. These mimic human touch, detecting things like pressure, temperature, and humidity. Fingertip sensors are particularly important because they are a key step for commercializing dexterous robotics hands, something policymakers have prioritized (NCSTI, November 3, 2023). U.S. firms currently lead in this segment. But an industry is emerging in the PRC. Key players like Suzhou Leanstar (苏州能斯达电子), Beijing Tashan Technology (北京他山科技), PaXini Tech (帕西尼), and Moxian Technology (墨现科技) all produce cutting-edge tactile sensors, AI-powered touch chips, and multi-dimensional force detection technologies. They also have secured funding to fuel further growth (China Daily, January 9; Xinhua, March 28; Sina, July 17; Sina, September 25).

Figure 1: The PRC’s Robotics Industry Supply Chain

 

 

 

 

 

 

 

 

 

 

 

(Source: Author research)

PRC Firms Begin to Flex Robotic Muscles

Humanoid robots depend on a device known as an actuator to mimic the movement of human muscles. Actuators are effectively the “muscle system,” whose component parts ensure that a robot’s joint function correctly. Contained within an actuator is a motor, which converts energy into motion, and a reducer, which works similarly to a gearbox to reduce the speed of the motor and increase torque. An additional critical feature is the ball screw, which employs ball bearings to facilitate rotational movement. The PRC has spent several years building its domestic industry across the value chain for these components. It has made progress in many areas, but in others remains reliant on external suppliers, either for the components themselves or for the machine tools to manufacture them. [1]

Motors

Most humanoid robots use two kinds of motors: coreless motors and permanent-magnet synchronous motors (PMSMs). Coreless motors are usually found in fingers, eyes, and lightweight grippers, as they are suited for delicate tasks requiring minimal vibration. For instance, Tesla’s Optimus has 12 coreless motors in each hand (Lammotor, August 15, 2024). PMSMs are used in knee, hip, and shoulder joints. They are larger than coreless motors, but provide strong, smooth motion that more effectively mimics muscle-like movement. They also are more energy efficient than coreless motors.

The National Robot Standardization Technical Committee (全国机器人标准化技术委员会) is led by experts from the Chinese Academy of Engineering and tech giants such as UBTech CTO Xiong Youjun (熊友军). At the end of 2024, it released the 2024 Humanoid Robot Standardization White Paper. This highlighted coreless motors as a critical area to master, in part to reduce dependence on foreign suppliers. [2] Current global leaders include the Swiss firm Maxon, Faulhaber in Germany, and Portescap in the United States. Together, these three firms comprise more than 70 percent of global market share (Baogao Dating, January 3). For frameless torque motors, top players include U.S. firms Kollmorgen and Moog, and German firm TQ-Group. PRC firms now support over half of the domestic market, frequently offering lower prices than foreign counterparts. These include coreless motor producers like MOONS’ (鸣志电器), DINGS’ (鼎智科技), and VEICHI (伟创电气), and frameless torque motor manufacturers like HeTai Motors (合泰电机), Kinco Automation (步科自动化), Han’s Motor (大族电机), and Haozhi (昊志机电) (Caifuhao, March 12; Baijiahao, March 13).

Reducers

Humanoid robots usually contain 20–30 precision reducers (Arxiv, February 28, 2024; HDS, accessed September 30). These include harmonic reducers, which are often found in arm, finger, and neck joints and ensure that joint movements are smooth and precise, and rotary vector (RV) reducers, which instead are found in leg, hip, and shoulder joints, and are designed for heavy lifting and to have long lifespans.

Reducers are one of a number of target technologies for indigenization, according to the Ministry of Industry and Information Technology (MIIT) (NCSTI, November 3, 2023). The 2024 National Robot Standardization Technical Committee white paper also acknowledged that RV reducers remain reliant on foreign suppliers, with significant growth potential for domestic players to close the gap. This focus is based on an assessment of reducers’ importance to humanoid robotics, but also on an acknowledgment that the PRC remains heavily reliant on foreign suppliers. Globally, Japanese firms Harmonic Drive Systems and Nabtesco dominate the market for harmonic and RV reducers, respectively. Harmonic Drive Systems held around 85 percent of global market share in 2023 (HDIN Research, April 18, 2024). Nabtesco, meanwhile, claims that it provides around 60 percent of the world’s reducers for medium-to-large robot joints (Nabtesco, accessed September 30).

The PRC has already reduced its external dependence on harmonic reducers. In 2023, Leaderdrive (绿的谐波) captured 15 percent of global market share and 26 percent of domestic share for its products. By comparison, Harmonic Drive Systems accounted for 38 percent of the PRC’s domestic market in 2022 (Weibo, February 27, 2024). Smaller players occupy 1–5 percent of the global market (Caifuhao, June 6). [3]

Attempts localize production of RV reducers, however, remains a long way from achieving full import substitution. One firm, Beijing Chietom Precision Transmission Technology (Chietom; 北京智同精密传动科技/智同科技), has received government backing to scale production and establish humanoid-specific bases. The company’s chief scientist, Zhang Yueming (张跃明), is optimistic, noting that over 80 percent of RV reducers are manufactured within the PRC. However, the country remains almost completely dependent on foreign machine tools, importing around 90 percent of equipment, mostly from Japan (36kr, January 17). Another domestic player, Zhejiang Shuanghuan Company (浙江双环传动), leads an MIIT project to localize both RV and harmonic reducers. It now supplies reducers to Tesla’s Optimus robot and Unitree’s G1-series robots. It is steadily building market share at a time when foreign firms, such as Nabtesco, are seeing their market share decline (163, November 27, 2024; Caifuhao, September 18).

Figure 2: The PRC Robotics’ Industry Depends on Nvidia’s Physical AI Stack

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

(Source: Author research)

Software and Data Chokepoints Remain

The hardware surveyed above provides the brawn for humanoid robots. Beijing is also keen to indigenize production of software, which provides robots’ brains. Large language models (LLMs) and multimodal artificial intelligence (AI) play two roles here. First, they provide task understanding and planning. LLMs parse human instructions and translate them into structured goals. Vision-language models (VLMs) then align those goals with what the robot sees. And visual navigation models (VNMs) collect 3D environmental data to generate an optimal path for acting on the original instructions. Second, LLMs can generate motion-control code, producing the low-level commands that drive actuators. Compared with other kinds of rules-based or reinforcement-learning–based controllers, LLMs are much better at interpreting intent. However, they also require much higher computing power and energy demand. A compromise solution for humanoid robots is edge-cloud collaboration, whereby only large-scale reasoning is offloaded to run on the cloud.

Tasks like grasping objects requires integrating and analyzing data from a range of inputs. Tactile and visual sensors all provide information that can be used to train vision–touch–language–action (VTLA) models for these purposes (Baijiahao, November 1, 2024). In September, three Hubei-based robotics organizations established the PRC’s first open-architecture, large-scale multimodal tactile data engine and cognitive training center. The center aims to address tactile data gaps and tackle long-standing challenges caused by fragmented and low-quality data (Baijiahao, September 8).

Paucity of data, especially high-quality data, remains a problem (MIT, April 30, 2024). Data for robotics differ fundamentally from those used in AI language models, comprising three-dimensional, multimodal, and bionic perception inputs to mimic human experience and interaction. Collecting this data is often costly and slow, requiring humans to wear motion-tracking gear or sensor gloves to repeat the same actions thousands of times. Simulation increasingly fills this gap, allowing virtual experiments to run in parallel. PRC institutions are pushing to generate, standardize training data sets and pipelines for robotics. In 2024, Agibot, together with top national research institutes, unveiled the world’s first open-source real-world humanoid robotics dataset, encompassing over 80 categories of everyday skills (Baijiahao, December 30, 2024). And this summer, the largest embodied intelligent robot data collection factory was built in Tianjin. Spanning 130,000 square feet, it is the world’s largest hub for embodied intelligent data collection and model training. Its core focus includes tactile sensor development, humanoid robot manufacturing, and the production of high-precision data sets (People’s Daily, June 24).

Currently, Nvidia is leading the charge in building open-source models, such as Isaac GR00T N1, and platform, such as Cosmos, for robotics development. The U.S. firm is also collaborating with Google DeepMind and Disney Research on Newton, a GPU-accelerated, differentiable physics engine that enables robots to learn how to handle complex tasks (Nvidia, March 18). Nvidia’s platform is rapidly reducing training times. Earlier this year, it enabled the GR00T N1.5 model to be updated in just 36 hours. With traditional data collection methods, this would have taken several months (Nvidia, May 18).

PRC robotics developers are increasingly turning to Nvidia’s physical AI stack to power their own products (EET, May 13). This obviates the need for Chinese developers to reinvent core infrastructure, accelerating the nation’s advance towards embodied intelligence. At CES 2025, one of the world’s largest trade shows for new technologies, Nvidia CEO Jensen Huang highlighted 14 leading robotics companies that use its Cosmos platform. Of these, six were from the PRC (Global Times, January 9; Nvidia, accessed September 30). [4]

At the cloud layer, Alibaba Cloud and Sharetronic Data (协创数据) use GR00T for model training. And before its addition to the U.S. government’s entity list, Inspur (浪潮信息) provided Nvidia with digital infrastructure to build a simulation model (GMW, August 24, 2022; Caifuhao, August 26; Baijiahao, September 25). At the application layer, leading humanoid developers—Unitree, Fourier, AgiBot, UBTECH, Galbot, and the Beijing Humanoid Robotics Innovation Center—also use Nvidia models for training purposes (Nvidia, June 2, 2024, August 9; AgiBot, accessed September 30). Even components firms such as Orbbec (奥比中光), BYD Electronics, and Apuqi (阿普奇) use these tools for sensing and control (Nvidia, August 9; Sina, October 16; Apuqi, accessed October 16).

Conclusion

Robotics competition is a key part of U.S.-PRC technological competition. The PRC’s robot supply chain is advancing rapidly, with domestic strengths in cost-effective sensors, reducers, and motors. These are frequently driven by central and local policy support in terms of subsidies and other incentives. But chokepoints persist. The PRC remains reliant on external partners for high-precision manufacturing equipment for high quality RV reducers, advanced coreless motor technology, high end ball screws, and software and simulation tools for AI training. In key parts of the value chain for humanoid robotics components, firms in the United States, Europe, and Japan remain clear leaders. This lead is unlikely to remain over the medium-to-long term, however, without policies that restrict PRC access to leading-edge technology and reduce collaboration in this critical sector.

Notes

[1] There are several different kinds of ball screw. In general, the high-end market in the PRC remains dominated by Japanese and European players, with Germany’s Schaeffler and Japan’s THK and NSK together constituting 90 percent of market share (Sina, January 12). In 2023, four out of five planetary roller screws—a specific type of ball screw—sold in the PRC were imported. But even here, domestic competitors are emerging, such as Nanjing Technology and Equipment Manufacturing (南京工艺), Best Precision (博特精工), Hanjiang Machine Tool (汉江机床) and Shuanglin (双林份).

[2] Humanoid Robot Standardization White Paper 2024 (人形机器人标准化白皮书2024), National Robot Standardization Technical Committee, December 2024. Available on Scribd here: https://www.scribd.com/document/864882725/%E4%BA%BA%E5%BD%A2%E6%9C%BA%E5%99%A8%E4%BA%BA%E6%A0%87%E5%87%86%E5%8C%96%E7%99%BD%E7%9A%AE%E4%B9%A6-2.

[3] These include Zhejiang Laifual Drive (浙江来福谐波传动), Han’s Laser (大族激光), Ningbo ZhongDa Leader Intelligent Transmission (宁波中大力德智能传动), TC Drive (同川精密), and CTKM Harmonic Drive (中技克美谐波传动).

[4] These included Star Era’s Star1, Zhiyuan’s Expedition A2, Fourier’s GR-2, Galaxy General’s G1, Unitree’s H1, and XPENG’s Iron.