In recent years, the humanoid robotics sector has become a battleground for both technology and capital, with numerous companies worldwide making significant investments, leading to rapid advancements in technology. On the evening of March 7, ZhiYuan’s CEO, Zhihui Jun, officially announced plans for a new product launch, coinciding with Xiaomi’s accelerated investments in robotics and the March 8 appearance of Optimus in Tesla’s latest promotional video. These events have sparked widespread attention both within and outside the industry, signaling that the humanoid robotics sector is reaching a new developmental peak.
ZhiYuan: A Leading Domestic Player, Advancing Both Production and Technology
Among China’s top-tier humanoid robot manufacturers, ZhiYuan has demonstrated impressive growth momentum. In August 2024, the company launched multiple robot models and swiftly moved toward mass production, establishing its first-phase factory in Lin’gang Fengxian. By January 2025, the 1,000th humanoid robot rolled off the production line, marking a major milestone. For the entire year, ZhiYuan plans to manufacture 4,500 units of its “YuanZheng” model and 200 units of the “LingXi” model, standing out among domestic robotics companies in terms of production capacity.
On the technological front, ZhiYuan remains committed to innovation. Since September 2024, it has introduced several groundbreaking technologies, including the AI+ robotics communication framework AimRT, embodied intelligence, the RoboDual system, and the large-scale simulation framework AgiBot Digital World. These developments have established a comprehensive and advanced technology ecosystem, continuously driving the intelligence of humanoid robots forward.
On March 11, 2025, ZhiYuan officially launched the LingXi X2, an all-purpose exploration robot that marks a significant leap forward in mobility, interaction, and operational capabilities. In terms of mobility, the LingXi X2 features a flexible material shell and 28 degrees of freedom, achieving unmatched agility through its cutting-edge core components and motion control algorithms. For interaction, it is equipped with the “Silicon Opto-Motion” multimodal interaction model, enabling it to accurately perceive human emotions and respond appropriately. In operational tasks, the QiYuan GO-1 model grants zero-shot generalization capabilities, allowing for multi-robot collaborative work, making it applicable across numerous industries.
Xiaomi: Cross-Sector Collaboration Accelerates Robotics Expansion
Xiaomi has placed robotics at the core of its strategic expansion, producing remarkable innovations. Xiaomi’s CyberDog series of quadruped robots (nicknamed “TieDan” or “Iron Egg”) has achieved major milestones. The first-generation CyberDog debuted in 2021, followed by CyberDog 2 in 2023, which featured a lightweight design, reducing weight from 14kg to 8.9kg, faster charging (down to 90 minutes), and an enhanced focus on intelligent pet companionship. CyberDog 2 can control smart home devices via voice commands and, powered by reinforcement learning algorithms, can even perform backflips and other acrobatic moves.
Xiaomi’s full-sized humanoid bionic robot, CyberOne (“TieDa”), also boasts impressive features. Standing 177cm tall and weighing 52kg, it has 21 degrees of freedom and a peak torque of 300N·m. Capable of recognizing 45 human emotions and 85 environmental semantics, it currently serves as a companion for home care applications, with potential expansion into industrial sectors.
Beyond humanoid robots, Xiaomi has developed a wide range of intelligent robots, including robot vacuums and interactive smart home robots. Its laser-guided robotic vacuum series, with 2,000Pa suction power and a low-noise operation of just 60 decibels, has gained a strong foothold in the smart cleaning market due to its affordability.
Xiaomi places high strategic importance on its robotics division, with Lei Jun personally overseeing increased investment efforts. The robotics team has grown to over 200 engineers, and in 2025, Xiaomi plans to mass-produce over 100 units of its next-generation humanoid robot, with a particular focus on dexterous hand development—several versions of which are currently in concurrent development. CyberOne is undergoing phased mass production at Xiaomi’s Yizhuang plant in Beijing, with an official production progress update expected in April.
Tesla: Industry Leadership and the Expanding Future of Optimus
Tesla has firmly maintained its position as the global leader in humanoid robotics. On March 8, 2025, Optimus made multiple appearances in Tesla’s latest promotional video, showcasing its practical capabilities in factory logistics, battery handling, and precision dexterous hand operations. These demonstrations highlighted its advanced technological edge and real-world utility.
The pre-production version of Optimus V1 is set for limited release this year, with several thousand units scheduled to be deployed in Tesla factories by late 2025. By 2026, Tesla aims to scale production tenfold to reach 50,000–100,000 units, and by 2027, another tenfold increase is expected, significantly expanding its market reach.
As Optimus continues to advance, its supply chain ecosystem has also gained industry attention. Major automotive and robotics suppliers such as Bosch and Schaeffler play key roles in supporting Optimus’ production and iteration. Additionally, companies like Sanlian Forging and Xiamen Precision are becoming increasingly involved in Tesla’s humanoid robot supply chain.
Key Technological Focus: Exploring Breakthroughs in Embodied AI
One of the most critical areas in humanoid robotics is the development of embodied intelligence large models (VLA), which enable robots to understand their environment, process language instructions, and execute physical actions. The current VLA models are primarily categorized into layered models and end-to-end models.
The layered model approach divides tasks into three distinct processes—input comprehension, reasoning and decision-making, and action execution—each handled by a separate model. This method accelerates commercialization, making it the preferred choice for most emerging humanoid robotics startups in China.
In contrast, the end-to-end model integrates all three steps within a single AI model. While this method requires extensive training data, is difficult to develop, and remains far from commercial deployment, it is considered a necessary step toward achieving advanced embodied intelligence. Only a few companies, including Tesla and Stardust Epoch, insist on adopting this more challenging path.
However, training high-performance end-to-end embodied AI models presents significant challenges—the data bottleneck being the most pressing issue. Firstly, the amount of training data available for robotic applications is 100 times smaller than that used for vision-language models (VLMs), severely limiting scalability. Secondly, data acquisition remains highly difficult and expensive.
Collecting real-world robot operation data is prohibitively costly, while synthetic data lacks real-world effectiveness. Even using human movement data for training is still in the early experimental phase. Moreover, human-collected data may carry biases or “toxic” elements, and given that the hardware design of humanoid robots is still evolving, existing datasets cannot be easily reused across different models. These factors significantly constrain the progress of end-to-end embodied intelligence models.
Currently, no industry-wide solution has been found to resolve these issues. Instead, companies are racing to collect and refine their own datasets, attempting to first achieve successful model generalization in single-task applications before scaling up.
[Disclaimer]: The above content reflects analysis of publicly available information, expert insights, and BCC research. It does not constitute investment advice. BCC is not responsible for any losses resulting from reliance on the views expressed herein. Investors should exercise caution.