UBS invited four humanoid robot experts to share their industry insights, with experts coming from robotics research, Qianxun Technology, Galaxy Universal, and other humanoid enterprises.

Experts generally agree that the hardware of robots is developing rapidly, while there is still no consensus in the industry regarding AI models and data paths. Once AI models mature, experts believe that the demand for humanoid robots will be rapidly released, guiding hardware cost reduction and widespread application.

Currently, due to technological limitations, humanoid robots may first be applied in data collection centers. Regarding recent events such as the World Robot Conference and the Humanoid Robot Games, experts noted that the industry is developing rapidly, and the ecosystem is continuously expanding, but there is also homogenized competition. In the long term, it is necessary to establish long-term barriers through AI models and data.

Experts are optimistic about manufacturers with AI large model capabilities and leading core component companies, but are not optimistic about manufacturers that focus solely on hardware. They remain cautiously optimistic about the development of humanoid robots over the next five years.

01

AI Models and Data are Bottlenecks, and Paths Have Not Converged

Experts generally believe that AI models and data are bottlenecks in humanoid development, but the development paths of both have not converged.In terms of data, humanoid manufacturers use varying proportions of real data and simulation data for model training. Some experts believe that simulation data is limited by the errors in mapping to the real world and that the modeling cost is high, preferring to accumulate high-quality real data. Experts believe that an operator can produce 1,000 hours of data in a year, and 100 hours of real data can complete training for a single task. If simulation data technology matures in the future, experts believe that the accumulated real data and AI models can be reused and will not be overturned. On the other hand, companies like Galaxy Universal mainly use simulation data, supplemented by a small amount of real data to complete the “last mile” of model enhancement.Regarding AI models, although the ultimate route cannot be discerned, mainstream solutions in the short term still involve VLA, with various paths: hierarchical models, stacked reinforcement learning, and stacked diffusion models (e.g., world models). Experts believe that in the short term, there may be third-party companies developing commercial models for data monetization, but humanoid manufacturers need to have their own AI models and data to establish long-term barriers. At the same time, due to the current difficulty in applying models and data across different machine types, experts believe that companies that make progress in AI models first will gain a first-mover advantage, driving the commercialization of humanoid robots.

02

Hardware Development is Rapid, but There are Still Localized Pain Points

Experts believe that hardware solutions are continuously maturing, although improvements are still needed in power density, heat dissipation, and lightweight design.Experts believe that manufacturers focusing solely on the body will find it difficult to establish barriers, and competition is beginning to intensify, with leading companies like Yushu already having certain advantages in hardware and motion control. In the future, as humanoid applications industrialize, experts believe that hardware costs will rapidly decrease, and the performance requirements for hardware in different application scenarios will become clearer.Currently, many body manufacturers independently develop/design core components on the hardware side and then collaborate with component suppliers for production, with suppliers mentioned in the meeting including: reducers (Green Harmonic, Dazhu, Tongchuan), sensors (Xinjingcheng, Kunwei Technology, Landot Touch Control), motors (Fuxing Motor), dexterous hands (Yinshi Robotics), and robotic arms (Ruiliman), etc.In terms of main control chips and computing chips for robots, experts generally feedback that NVIDIA is the main supplier.

03

Scenes are Continuously Enriching, Data Collection Centers May Land First

Given that the technology for humanoid robots is not yet mature, experts believe that in the short term, humanoids will first be applied in data collection centers to enrich scene data and feed back into AI models.Experts mentioned that the government is supporting the construction of data collection centers through subsidies. At the same time, humanoid companies are continuously exploring and testing application scenarios. Galaxy Universal is collaborating with Tianqi Co., Bosch to explore scenarios in automotive factories; and with Meituan to achieve unmanned pharmacies through robot picking, aiming to land 100 stores by the end of the year. Qianxun Technology is using large models to accomplish tasks like folding clothes, and is optimistic about future application scenarios in healthcare, logistics, and to C.Currently, the efficiency, accuracy, and generalization of robots are the main issues hindering commercialization.