As 2026 reaches its halfway point, the humanoid robot and embodied intelligence industries have delivered several outstanding performance reports. In July, at the press conference for the World Artificial Intelligence Conference (WAIC), Gan Xiaobin, Deputy Director of the Department of Science and Technology at the Ministry of Industry and Information Technology (MIIT), revealed that the annual production of complete humanoid robots this year is expected to exceed 100,000 units. In the same month, at a press conference held by the State Council Information Office, Wang Weiming, Chief Engineer of the MIIT, stated that quadruped robots developed in China account for nearly 70% of global sales, and there are over 400 models of complete humanoid robot products, accounting for more than half of the global total.
2026 is widely regarded by the industry as the first year of large-scale commercialization for humanoid robots and embodied intelligence. Meanwhile, the evaluation metrics for the entire industry are undergoing a fundamental shift—from whether a prototype can be built to whether actual benefits can be created. In this process, the embodied brain, centered on large models, is becoming a key variable in defining the core competitiveness of the industry.
Why Is the Embodied Brain the Most "Lucrative"?
Entertainment performances, intelligent industrial manufacturing, education and scientific research, commercial retail, home services... Today, an increasing number of robots are stepping out of industrial assembly lines and entering broader scenarios, facing entirely new technical challenges. Previously, traditional industrial robots mostly operated in closed, structured environments. Facing single and deterministic tasks, they could complete them without environmental perception and autonomous learning capabilities.
However, in real-world scenarios with infinite possibilities, the same object may appear in different locations, and different objects may be placed in the same location. Different lighting conditions and different materials will affect recognition effects, and obstacles may also appear during task execution. What machines need more is the learning ability to understand the laws of the physical world and the generalization ability to solve complex problems.
This is exactly why embodied intelligence large models are so critical. Gao Jiyang, Founder and CEO of Xinghaitu, told reporters from China Electronics News that the greatest opportunity for the embodied intelligent robot industry to realize productivity lies in the fundamental upgrade of the intelligence paradigm. Hardware is the carrier, while intelligence is the soul. Only by continuously evolving the "brain" can the hardware release its true industrial value. He noted that the market focus is shifting from "stacking hardware and competing on parameters" to "competing on models, algorithms, and scenario generalization capabilities."
Amid the cycle shift, the long-term value of the "brain" layer is being systematically re-evaluated, and the investment focus in the primary market continues to tilt from the "body" to the "brain." According to statistics from relevant media, in the first half of 2026, financing in the core track of domestic embodied intelligence reached 43.8 billion RMB. Financing for companies focusing on the embodied brain accounted for over half, while financing for companies centered on the robot body accounted for only 12.8% during the same period. Embodied brain companies are accelerating capital aggregation. Tashi Zhihang, an embodied intelligence startup founded two years ago, announced in April the completion of a 455 million USD (approximately 3.3 billion RMB) Pre-A round of financing, setting a new record for the highest single-round financing in the domestic embodied intelligence field. Similarly, Ant LingWave, founded just over a year ago, hardly produces hardware and focuses on the embodied intelligence large model track, and has recently confirmed the launch of its first round of independent financing.
A strategic consensus across the industry to step up investment in the "embodied brain" has been formed. Embodied intelligence companies queuing for IPOs have also demonstrated their determination to engage in large model R&D. Recently, the official website of the Shanghai Stock Exchange disclosed the reply to the first round of audit inquiry letters for Deep Robotics' STAR Market IPO. It mentioned that the raised funds will focus on frontier directions such as the "embodied brain" and "embodied cerebellum," with about 90% of the funds allocated to "brain"-related technologies and infrastructure to make up for previously relatively weak resource investments. Unitree Robotics, which went public earlier, also allocated up to 48.13%, or over 2 billion RMB, of its IPO raised funds to R&D in embodied intelligence algorithms and large models.
Zhang Dapeng, Assistant Vice President of Leju Robotics, stated that full-stack in-house R&D capabilities have become a watershed in industry competition. Companies with fully autonomous and controllable capabilities across the entire chain of "brain-cerebellum-body-data" demonstrate stronger advantages in large-scale implementation.
How Should the Next-Generation Embodied Brain Evolve?
Embodied intelligence large models have an essential difference from traditional pure language large models. Zhang Dapeng emphasized that embodied intelligence requires "embodied data"—that is, high-quality operational data generated when robots execute tasks in real physical environments. This difference determines that the core mission of embodied large models goes beyond processing semantics; it also requires understanding the physical world and making independent decisions.
So, which route is most likely to lead to the future of the embodied brain? If this question were asked a few years ago, the answer would almost undoubtedly be the VLA (Vision-Language-Action) model. By integrating these three types of models, robots can observe the environment, understand instructions, and then translate that understanding into executable actions.
However, with the advancement of robot technology, the limitations of the pure VLA route are gradually being exposed. For example, Yao Maoqing, President of Agibot's Embodied Business, believes that VLA models have not learned the representation of universal physical laws; changing the robot body or facing an unfamiliar scene leads to an immediate decline in generalization. In addition, data scarcity and insufficient stability are also challenges that cannot be ignored.
Since the beginning of this year, technical directions such as world models and WAM (World-Action Model) have attracted much attention. Unlike VLA models, which excel in natural language understanding, world models are better at prediction, helping robots learn the laws of the physical world and deduce the results of action interactions. The WAM model goes a step further to unify the two major technical paths of VLA and world models, possessing both the ability to execute actions and predict the world.
Different technical directions are not mutually exclusive options. Currently, the parallel development of VLA models and world models is becoming a common choice for many leading enterprises. Unitree Robotics released the WVLA 2.0 embodied large model, integrating world models and vision-language-action models; Galbot launched the general brain foundation model AstraBrain WAM 0.5, with its core technical route being the world-action model; Xinghaitu implements a "one brain, multiple forms" strategy, releasing the new-generation VLA foundation model G0.5 and the Fast-WAM world model; Agibot is also simultaneously pushing forward on the dual tracks of the GO-2 embodied foundation large model and the GE 2.0 world model.
Notably, within the world model track, the technical focus of different enterprises also shows a clear divergence. At this year's WAIC, six enterprises—Daxiao Robotics, Ant LingWave, Extreme Vision, Boundless Dynamics, Agibot, and Independent Variable Robotics—communicated on the same stage. Xia Zhongpu, Co-founder and CTO of Boundless Dynamics, stated that evaluating a world model depends on its prediction accuracy for geometry, motion, and force. Shen Yujun, Chief Scientist of Ant LingWave, believes that physical rationality is more important than accuracy. Ren Guanghui, AI Algorithm Director at Agibot, regards "long-term physical consistency of reasoning" as an indispensable evaluation metric; Wang Hao, Co-founder and CTO of Independent Variable, emphasized the importance of the decision window period, believing that the timeliness of reasoning is more critical.
Although the technical routes vary, as Tao Dacheng, Chief Scientist of Daxiao Robotics, said: "Divergence is the material for papers, while overlap is the foundation of the industry." When robots truly enter the physical world, the evaluation standards converge on the same goal: whether they can stably complete tasks. The truly valuable opportunities may lie hidden in those paths where consensus has not yet been reached.
How to Overcome the Data Threshold?
However, whether it is the VLA model or the world model, neither can bypass the same core crux—data.
In Zhang Dapeng's view, data is the most easily overlooked but most likely to become a long-term bottleneck in the industry. Compared with text data, the collection cost of embodied data is higher, the standardization difficulty is greater, and the requirement for scenario coverage is broader, making it much harder to obtain than internet text data. Yao Maoqing stated that current embodied data, whether calculated by tokens or duration, may differ by more than 10,000 times compared to the training data of leading large language models. If the robot "ChatGPT moment" is defined as robots being able to understand open-ended instructions, be "out-of-the-box" ready in the physical world, and complete common tasks, the industry may need to accumulate data at the level of 100 million hours to push the basic success rate of common tasks to 70% to 80%.
At the same time, data type mismatch is also an important factor restricting the development of the embodied brain. Taking the world model as an example, Yao Maoqing believes that there is an essential difference between the data distribution of internet videos and the physical world data that robots truly need. Much of the internet content is entertaining or even contrary to physical laws, whereas robots need real data rich in physical contact and covering various types of physical interactions. These are not only scarce on the internet but also difficult to simulate through existing physical systems. Huang Tiejun, Chairman of the Beijing Academy of Artificial Intelligence (BAAI) and Professor at the School of Computer Science at Peking University, also believes that in the world model stage, an increasing amount of real-time interactive data will be needed, and wearable devices such as earphones and smart glasses will accelerate this change.
To solve this problem, the industry is exploring feasible paths for data acquisition. Galbot is betting on the data synthesis track, making up for the shortcomings of real data in simulation environments. Xinghaitu, together with Yizhuang Robotics and Yizhuang State-owned Capital Investment, jointly established the data company "Yishu Intelligence," planning to complete the accumulation of one million hours of data in the current year and scale up to tens of millions of hours within three years. Leju Robotics is taking the lead in building an open-source dataset for embodied intelligence, enabling research teams and small and medium-sized robot enterprises to use it "out-of-the-box."
Thus, it can be seen that technical bottlenecks are both challenges and springboards for industrial upgrading. The turning point of the embodied intelligence industry is hidden in every set of data accumulation, every round of model iteration, and every product landing. This industrial competition concerning the next generation of productivity is not a sprint that happens overnight, but a long-distance run that requires patience and perseverance.
Author | Chen Cun, Editor | Qiu Jiangyong, Art Editor | Ma Liya, Supervisor | Lian Xiaodong