Yushou's robots have tested two "brains" in just two weeks. Source: First Financial. On October 12, Wang Xingxing stated at the 26th China International Industrial Expo that the release of GPT-6 has profound implications for the embodied intelligence industry, and the company's team has deployed it in robots for task testing, achieving quite good results. He also mentioned, "Currently, my personal faith in AI is actually increasing day by day. The biggest mistake I made in the past decade was not trusting AI enough. To this day, many people may not have fully utilized or embraced AI."
On October 15, Yushou's stock closed at 405.10 yuan, marking the lowest price since its listing two months ago, with a market value of 163.8 billion. On October 19, NVIDIA's DreamZero team announced the world's first publicly available robot brain API, featuring the team's first model, Dream-1. After testing 27 robots, including Yushou R1, these robots were able to plug and play without prior training, achieving tasks such as folding towels, opening drawers, hanging headphones on hooks, and "placing non-kiwi fruits into a basket."
On the same afternoon, Yushou's market value returned to 200 billion. In just four days, Yushou's market value rose over 22% from its low point, reaching 208.3 billion at today's close, remaining above 200 billion. This raises curiosity about what changes have occurred. Along with the increase in market value, two tests provided judgments: embodied intelligence is no longer just a possibility.
The model has crossed the threshold of embodied capabilities. When ChatGPT-6-Astra was released, it stated, "Welcome to the AGI era," claiming that the release of this model marks the arrival of the AGI era. OpenAI referred to it as "the most intelligent model in the world that aligns with human intentions." Since it claims to understand human intentions, how well does it actually understand? Can it comprehend the physical world like a human? To further understand whether GPT-6-Astra possesses the ability to control embodied intelligence, a research team from SenseTime, Daxiao Robotics, Ropedia, Nanyang Technological University in Singapore, and the University of Tokyo conducted tests involving multiple real machines, including the Yushou G1 humanoid robot.
SenseTime's SenseNova and its joint research team deployed it across eight real robot platforms, covering five different forms, without providing any demonstrations to GPT-6-Astra. The model only received structured information written by humans, including robot interfaces, joint constraints, camera positions, safety constraints, and operational prompts on some platforms. Under these conditions, the same GPT-6-Astra was indeed able to control robots of different forms to complete tasks involving rigid object grasping and placement: it succeeded 17 out of 18 tests.
On the Yushou G1, GPT-6-Astra completed 2 out of 3 robotic arm operations through the underlying control interface, as well as all 3 tasks of turning, walking, and stopping.
Wang Xingxing emphasized "generalization" at the WRC (World Robot Conference) in August, hoping to promote technology to situations with better generalization capabilities. From this test, it can be seen that GPT-6's adaptability on multi-form and multi-platform robots has completely surpassed the level of two months ago. However, merely grasping and walking is not enough for robots to truly integrate into real life. The report concludes that GPT-6-Astra has largely crossed the threshold of embodied capabilities, but practical deployment of embodied intelligence is still a considerable distance away. Despite existing issues in practical applications, this test proves that embodied intelligence is no longer just a backdrop for storytelling, but a concept that is truly becoming a reality.
Success of GPT-6-Astra in Embodied Intelligence Testing
Wang Xingxing also stated at the WRC, "Currently, the biggest bottleneck in global embodied intelligence is the degree to which AI model inputs and outputs align with robots. It can execute your tasks, but the final few millimeters of deviation cannot match the real world well." This is a part that language models still need to resolve. "For example, when I reach for something, I see that I'm almost grabbing it, but the last bit of tactile feedback cannot be corrected well." He believes this issue will definitely be resolved in the coming years.
Meanwhile, specially trained foundational models for robots have emerged. On October 20, Dream Labs, founded by the DreamZero team, completed a $90 million seed round of financing and simultaneously launched Dream API: the world's first publicly available robot brain API. After registering, developers can connect their robots to the API for use. The model is deployed across multiple data centers globally, allowing for quick responses regardless of the robot's location. In terms of pricing, there is the Dream-1 at US$3 per hour (approximately HK$23), and a faster, lighter version, Dream-1-flash, at US$2 per hour (approximately HK$16).
However, Dream Labs not only aims to make it "usable by everyone" but also "usable by every robot." Dream-1 indeed possesses this capability. Its training does not rely on text or deployment on a specific machine but is based on videos of people or machines completing tasks. The model learns to predict both videos and actions simultaneously, thus making "controlling a robot" and "imagining what the robot will do next" the same problem, which is an advantage of physical AI models.
Therefore, for Dream-1, completing tasks is merely an expression of "objects," and all "objects" movements are a way to execute commands. The specific instructions for a particular machine are not important to it, eliminating the need for repetitive input of commands from the source. For example, in the case of folding clothes, a robot controlled by GPT-6 needs to have additional depth information to fold towels or tablecloths, which are flexible objects with consistent parameters, but the complete task of folding clothes can never be accomplished—completely different fabrics, shapes, wrinkles, and changes in center of gravity make it too difficult for the language model to judge accurately.
Dream-1 can now complete the instruction to fold clothes, and the official website lists folding tasks as one of its key capabilities, with the real machine able to fold towels in a zero-sample situation.
In the RoboDojo-RC Tier 1 evaluation released by the independent third-party evaluation agency Robocurve, Dream-1 achieved an average score of 59% per task attempt, becoming the first model to exceed 50%, scoring higher than Claude Opus 5.5 (48%), Gemini 4 Argon (48%), GPT-6 Astra (47%), GPT-6.1 Sol (42%), and other latest large models, as well as higher than MolmoAct2 (19%) and GR00T N1.7 (10%) robot models.
Besides the scores, Dream-1 is also cheaper and faster: the cost per task attempt is about US$0.15 (approximately HK$1), scoring 1.2 times higher than the second-highest scoring Gemini 4 Argon, but costing about 16 times less. The average time taken is 1.4 minutes, scoring 1.3 times higher than the fastest large model GPT-6 Astra (ultrafast) and being 1.7 times faster. Each model configuration has been tested 120 times in the real world.
This does not mean that Dream-1 is a fully mature robot model. The model only determines where the gripper ultimately goes and does not control how the arm reaches there, which may cause issues near obstacles or at joint limits; when facing a body that differs significantly from previously seen machines, it may predict infeasible actions; furthermore, there is still a gap between zero-sample and one-time adaptation. Dream Lab co-founder and CTO Ge Yunhao stated, "Narrowing this gap is one of our main goals. The stronger the zero-sample capability, the closer any new robot will be to being plug-and-play."
Regardless, as Ge Yunhao stated, "Just like when ChatGPT was first released to the public, robots have finally welcomed the first publicly available general brain API. It may not be perfect yet, but someone has to take the first step."
Yushou: Present at both events. In this experiment, Yushou's G1 and R1 were both present. This is no coincidence. Yushou robots are well-known, but why were they used in both experiments? What sets them apart from other robots? From an industry stability perspective, Yushou's revenue reached 1.152 billion yuan in the first half of 2026, a year-on-year increase of 48.54%; net profit attributable to the parent company was 274 million yuan. This profit figure stands out in the entire industry. During the same period, UBTECH reported a loss of 339 million yuan in the first half of the year, with cumulative losses exceeding 3.5 billion yuan since its listing; Zhiyuan reported a full-year loss of 83.54 million yuan in 2025, with revenue of 1.05 billion yuan.
The only company that can actually produce, sell, and still make a profit is Yushou.
Compared to other robots, Yushou also has a significant price advantage. According to Frost & Sullivan data, in the first half of 2026, Yushou ranked third in global humanoid robot revenue, with a market share of 15.3%. This data may seem contradictory to the previous text, but it is not. Companies with higher revenue have lower sales, which is not contradictory. The pricing of Yushou G1 starts at 99,000 yuan, making it one of the cheapest full-size humanoid robots in the industry; other manufacturers' industrial-grade humanoid robots generally range from 200,000 to 500,000 yuan. Thus, among robots of the same category, disregarding all revenue data, Yushou is also the most widely circulated, lowest-priced, and has greater short-term commercialization potential.
Once the embodied brain is truly realized, Yushou will undoubtedly be one of the top choices for physical hardware.
The turning point for embodied intelligence is now closer than ever. Returning to the increase in market value, why has Yushou been so significantly affected, and what is the secondary market betting on? Yushou's shipment volume has always been high, which is not a factor affecting the rise in Yushou's stock price. When ChatGPT-6 was released, JPMorgan still gave it a "reduce" rating for a straightforward reason: overvaluation and overstretched expectations. This also proves that there has not been a "brain" that the market finds convincing, thus robots have merely been tools that somewhat resemble humans, rather than embodiments of intelligence.
However, on the day Yushou's market value returned to 200 billion, GPT-6 and Dream-1, in the same month, provided the same result from two different directions: large models have developed to the point where they can control robots without relying on robot data. When all robots can connect to the same brain, the one with the largest hardware circulation in the market will become the default testing platform, the default data source, and the default implementation solution. Yushou's cumulative shipment of 18,000 units, starting price of 99,000 yuan per unit, and its unique profitable financial status create a moat for it to seize its ecological niche.
This elevates Yushou's status from being mocked as a robotic toy to becoming a stable carrier of embodied intelligence.
Initially, people believed that Yushou could independently achieve "embodied intelligence," sparking a frenzy; once calmed down, people thought it was unwise to place expectations solely on a hardware company. Now, embodied intelligence has provided the market with an answer: separation of head and body, integration of brain and machine. This also opens up much imagination. In the future, hardware will become cheaper, brains will become more numerous, and embodied intelligence is gradually stepping into human life.