The 2025 World Artificial Intelligence Conference (WAIC), which opened last weekend, was a spectacular event. Hundreds of top-tier Chinese AI companies gathered in Shanghai, offering the public a clearer glimpse into the future of AI.

On the morning of July 27, the Zhiqi Embodied AI Forum, hosted by Zhiyuan Robot and themed 'New Opportunities for Embodied AI from a Global Perspective,' was held on the sidelines of WAIC. Yao Maoqing, Partner and President of the Embodied Business Division at Zhiyuan Robot, revealed two major announcements during his keynote speech:

First, Zhiyuan Robot launched Genie Envisioner (hereinafter referred to as GE), the industry's first open-source platform for world models specifically designed for real-world dual-arm robots.

Second, Genie G2, the next-generation body of the Genie series, is scheduled for release in Q3 this year. It features comprehensive upgrades in motion accuracy and scenario adaptability, along with native data acquisition capabilities, further expanding the application boundaries of embodied AI.

What Is the Architecture of a World Model?

Yao Maoqing explained that the launch of GE actually encompasses four key initiatives, which can be intuitively visualized through an airplane graphic.

At the center is GE-Base. After being trained on massive amounts of real-world robot data, it achieves robust video generation capabilities, specifically excelling in long-range, multi-view video synthesis.

How does GE-Base achieve this 'two wings flying together' capability?

Zhiyuan developed a World Action Model through action generation training. Given specific instructions, the robot can anticipate future scenes and plan actions. By adding an Action decoder to the video generation pre-training framework, the model enables long-horizon task planning. Robots based on this model perform significantly better in dynamic tasks such as folding clothes and conveyor belt sorting.

The other component is the Action-conditioned World Model. Under given trajectory plans, it allows the robot to conduct high-fidelity physical simulations, accurately simulating subsequent environmental interactions and changes in the physical world. Zhiyuan conducted a comparison between the environment predicted by the world model and the actual real world, finding a very high degree of fidelity.

In quantitative evaluation, the success rate of robots in real-world tests closely matched their success rates when evaluated within the world model.

Finally, Zhiyuan also launched a Benchmark specifically for GE, designed to evaluate the rationality of generated content. Evaluation dimensions include the rationality of world model generation, motion self-consistency, semantic plausibility, spatiotemporal causality, and visual simulation accuracy. This Benchmark has currently been open-sourced on GitHub.

According to the plan, Zhiyuan will officially release GE-Base, World Action Model, and Action-conditioned World Model in the third quarter.

Why launch GE?

Yao Maoqing, partner at Zhiyuan, highlighted a critical point: the embodied AI flywheel, which consists of 'body-hardware - data - model - scenario'.

On the hardware side, robots need to possess high-quality, robust physical bodies; on the data side, the industry has embarked on a data-driven path, requiring the construction of massive datasets across various sectors; by leveraging this vast amount of data, models are continuously refined for application scenarios; meanwhile, during application processes, new requirements emerge for hardware design and data methodologies, continuously driving the flywheel and accelerating development.

This also led Yao Maoqing to remark that developing embodied AI is extremely challenging.

He stated: "Many times I envy friends working on large language models, as it seems they can complete all their work just by sitting in an office. But we are different; every day requires continuous self-evolution across hardware, data, models, and scenarios."

To establish the flywheel of Embodied AI, data is the fundamental prerequisite. Unlike the massive datasets available in the large language model sector, the robotics industry currently lacks sufficient data. Therefore, Zhiyuan has established its own specialized professional data acquisition factories, forming AgiBot World—the world's largest dataset—and open-sourcing it.

Yao Maoqing stated: "Facing the desert of Embodied AI data, we chose to plant the first tree, hoping it will eventually grow into a forest."

With data secured, the next step involves models. Prior to this, Zhiyuan released the industry's first general-purpose Embodied Foundation Model—the Qiyuan Large Model—which possesses 'one brain, multiple forms' capabilities, enabling adaptation to other heterogeneous robot bodies.

However, the emergence of physical intelligence has only just begun. Robots should not merely execute human commands; they must also be equipped with the wings of imagination to foresee the future. This vision served as the primary backdrop for the introduction of GE.

In terms of scenario implementation, Zhiyuan Robots have successfully deployed solutions across four major sectors: industrial manufacturing, warehousing and logistics, power inspection, and interactive guidance.

Thus, the Embodied AI flywheel for Zhiyuan Robots has been formed. Throughout its long-term R&D process, Zhiyuan encountered numerous challenges while achieving significant gains. Yao Maoqing shared four key insights derived from this journey.

He believes that robots must return to the essence of 'actuators.' High-frequency, high-precision tracking and low-latency actuators are crucial. Furthermore, high-precision and high-resolution force control are essential to enable operations in industrial scenarios.

At the sensor level, robots not only require multi-sensor fusion but also precise calibration, time synchronization, and consistency across multiple machines.

In terms of body configuration, no matter how powerful the algorithms and models are, they cannot replace the advantages of a high-degree-of-freedom flexible body. Zhiyuan hopes that in the future, robot bodies will become increasingly similar to human configurations and offer higher degrees of freedom.

Finally, there is the amplification effect of systematic errors. Whether in hardware manufacturing or model training, the accumulation of small errors can ultimately be amplified infinitely, posing a challenge that robots must overcome for large-scale applications.

This is precisely the summary from the scenario side, which will ultimately make Zhiyuan's next-generation product even better—the soon-to-be-released Genie G2.

Zhiyuan Robot, which has formed the embodied AI flywheel, may perhaps run faster in the implementation of various scenarios.

Appendix: Transcript of the Conversation

After the speech, Yao Maoqing, partner of Zhiyuan Robot and President of the Embodied AI Business Department, was interviewed by multiple industry media outlets including No. 42 Radio Waves, who conducted an in-depth discussion on Zhiyuan Robot's latest commercialization progress, world models, and other topics.

Q: How does Zhiyuan's world model differ from others?

A: From a technical perspective, there are two main approaches for world models. The first, which we might call the 'general VLA era,' adds a decoder on top of a Vision-Language Model (VLM) to directly generate actions. However, world models were originally not designed to generate actions but to simulate the world.

We need to model many aspects of the physical world, but explicit modeling is complex and limited. Consequently, people have turned to using neural networks to brute-force process data. Starting around 2022, Tesla and another autonomous driving company were working on this. Initially, their task was based on the car's front-view camera: telling the vehicle to turn forward, go straight, or avoid running red lights. Later, the task shifted to instructing it to grab a bottle of water. Instead of just issuing a command, the system provides control instructions at 50 Hz (every 20 milliseconds), specifying joint angles and a series of action controls that must be mimicked to produce a complete execution. This is what we refer to as an Action-conditioned World Model, where 'Action-conditioned' implies action-driven. The other category is the World Action Model, which has a conjugate relationship with the former.

One type takes given actions to generate future visual frames, while the other takes a desired final state to generate the necessary actions. For example, if I tell you to grab water, you not only visualize the scene but also generate the precise control signals for those seven joints at every moment, determining exactly what angles they should assume. This concept is similar to our earlier discussion about an airplane taking off with its wings.

Q: Some competitors rely on synthetic or virtual data. What is your view on these two types of data collection?

A: Our forum today included such discussions. Many companies tend to reverse-engineer this problem from their own starting points, aiming to provide answers like a 1:10 or 1:99 ratio of real-to-synthetic data usage.

In reality, it is difficult to cover all the tasks we want robots to perform in the physical world using only simulation. While local tasks might be handled via simulation, most reliance must still be placed on real-world data, which is also the effect we have observed in our experiments. Furthermore, among teams that claim to rely heavily on simulation, if you look deeper, you will find that they are recently collecting massive amounts of real-world data as well.

Q: You previously worked at EV companies like NIO. How do you view robots entering factories for practical training and starting to play a role? How much more time do you think is needed?

A: I am relatively optimistic. We have already signed projects for flexible scenarios earlier this year, and we expect to see many such cases emerge within the next year.

Q: There is still a gap between current capabilities and replacing human employees. What are the main difficulties? Which areas do you aim to breakthrough when collaborating with industry partners?

A: We have already achieved 10,000 consecutive zero-error operations, reaching human-level performance. However, upon actual deployment, expectations are higher—for example, production capacity must not be affected, which can also be managed through other operational measures.

Q: In factory settings, some robot manufacturers adopt battery swapping or wireless charging. What about your approach?

A: Our next-generation products will support both. They enable hot-swappable batteries without powering off, as well as autonomous recharging, similar to home robotic vacuums that charge automatically when the battery is low.

Q: We are quite curious about Zhiyuan's technical roadmap. What are the main challenges? Regarding the 'brain' and 'cerebellum' versus the body, where does your team focus its efforts?

A: As mentioned in our report, achieving deployment requires a closed-loop flywheel effect; all these aspects must be handled well. It is difficult to completely outsource any single part. In practice, I believe it is an iterative process—a cycle of designing something at a certain point, collecting data afterward for iteration, discovering new areas needing improvement, and then revisiting the body design. Therefore, regarding resource allocation, our team continues to work hard across all directions.

Q: In early July, Zhiyuan won a 78 million RMB order from China Mobile. What are the main components of this order, including product unit prices, application scenarios, and specific functionalities?

A: This project involves a custom-built bipedal humanoid robot for China Mobile. The product is primarily designed for use in telecom operators' retail stores and business halls to perform tasks such as reception, guided tours, and interactive services.

Q: Will you continue to focus on the service robot sector next?

A: Yes, we will. This initiative marks our first step into similar scenarios. In the future, we expect to see service and reception applications deployed in operator-managed hotels, banks, and other venues, where there is massive demand.

Q: Can we understand that service reception or guided tour scenarios are your primary direction? Are there other application scenarios planned next?

A: This is just one of our application methods. By leveraging AI agents to create infinite capabilities and enabling interaction, we are bringing robots into factories. Factory deployment is only the first step; the next phase targets the retail and service industries, with the ultimate goal of entering households in several years.

Q: Many investors believe that the window period for Embodied AI lies between the second half of this year and the first half of next year. When do you think Zhiyuan will deliver a significant result to the market?

A: We will submit our 'report card' in the second half of this year. I believe this assessment is accurate. Many teams have emerged recently, producing similar PR materials and demos—whether for factory work or home assistance. It is difficult to determine whose capabilities are truly superior based solely on these demonstrations. The most critical method for verification is practical, real-world testing.

Many leading companies, such as Zhiyuan, have reached a relatively high level. At this stage, everyone looks at your revenue and performance. In the second half of the year, we will enter a phase where we deliver tangible results to industries with increasing intensity.

Q: Will there be homogenization?

A: There will definitely be some homogenization. We see that after many directions we’ve released are announced, numerous competitors tend to follow. However, I believe the market is still large enough, and the cake is big enough, so currently several top players all have opportunities.

Q: Will Zhiyuan adopt an Android-like model, opening up hardware interfaces and focusing on model licensing?

A: First, from a business model perspective, I personally feel we won’t follow the Android approach. Android belongs to Google, whose core business model is that of an internet company, primarily relying on internet traffic, advertising, search, and other services—with Android mainly driving search revenue.

As a hardware manufacturer, Zhiyuan provides more integrated software-hardware solutions, offering aggregated optimizations to our customers. This is more akin to Apple or Tesla, which optimize across the full stack of both software and hardware, delivering a seamless user experience as seen in smartphones. I hope our products can bring users an ultimate experience.

Q: How do you respond to automakers entering the field, for example, XPeng?

A: Automakers entering the robotics space is undoubtedly an inevitable process. They possess their own advantages in many areas, such as supply chain management, organizational efficiency, smart manufacturing capabilities, and accumulated expertise in autonomous driving.

I come from the autonomous driving industry. We believe that automakers may not be as focused as Zhiyuan, which is solely dedicated to this field. For us, this is the only thing we must do, and we have no choice but to excel at it.

In larger companies, whether automakers or internet firms, they are mostly in the early stages of layout because this is not their core business. No matter how well you perform, from a financial report perspective, it might primarily result in R&D expense losses. Their current investments are relatively limited, though we cannot rule out them stepping on the gas pedal once our path is proven.

Compared to industries like new energy vehicles, the addressable market for embodied AI is much larger. Products in the NEV sector are highly homogenized—essentially four wheels plus a sofa. In contrast, robots can be applied across many industries, with specialized companies emerging in each sector and niche market.

Q: Taking one of Zhiyuan's products as an example, discuss the current progress of the industrial chain, particularly regarding improvements in product delivery speed.

A: This year, we have entered a concentrated phase of commercial interaction, with shipments expected to reach several thousand units. Currently, the supply chain indeed poses a significant challenge.

Achieving large-scale capacity while ensuring product consistency remains a process of overall improvement. At this stage, the suppliers we engage with are growing alongside us; most are still small to medium-sized enterprises. Previously, major players in electromechanical components were hesitant but are now eager to participate. Our recent ecosystem partners are preparing to enter the field after seeing the greater opportunities presented by our vehicles.

Q: What has been Zhiyuan's biggest breakthrough over the past year, specifically in which area?

A: Our joint modules, especially those used in our bipedal humanoid robots, went through mass production ramp-up last year and have basically stabilized this year. Last year's status was difficult; for instance, Chery, known for its 4S shop robot, inspected our products using automotive industry standards, which was quite painful. In the early stages of the industry, it took us a long time to reach such high standards. However, we have found that once we overcome this hurdle, it represents a massive transformation for us.

Q: Zhiyuan has demonstrated capabilities across various scenarios. Do you believe we have now reached a stage of systematization, or are companies still largely learning through trial and error?

A: There are differences among them, as they are all startups, which makes the challenge inherently difficult. Focusing on specific scenarios is a common approach; some target retail, others industrial applications, and some home environments. I consider this to be normal.

Even for Zhiyuan, we do not cover every industry scenario. Through extensive long-term screening, evaluation, and client recommendations, we have identified scenarios that align with our technology. Our primary operational scenarios are mainly in manufacturing and logistics.

Q: Regarding implementation, how deeply are you involved in setting standards?

A: We are currently in the early stages. We handle end-to-end development and interaction design ourselves. This includes hardware improvements, model calculations, and final engineering deployment—all areas where we participate personally.

Q: Will this change in the future?

A: It certainly will. The goal is to encapsulate this development system into a robust framework, empowering downstream ecosystem partners to complete certain scenarios with low barriers to entry and facilitate low-threshold delivery. This is what we aim to achieve.

Q: Have you noticed any shifts during communications with application scenario providers? Has their acceptance level changed during industry exchanges?

A: As embodied AI becomes more popular among the public, many application-side stakeholders have undergone a cognitive shift. Last year, when we interacted with clients, they viewed it as an automated ecosystem—something you could simply plug in and use directly.

Today, they realize that this requires mutual effort from both sides; it is a data-driven process that necessitates training robots in real-world scenarios for continuous iteration. Additionally, production line modifications are needed to facilitate better robot integration. From a financial perspective, clients are willing to invest significant funds in conducting POC (Proof of Concept) validations with us, rather than adopting a 'ready-to-use' approach.

Q: From an entrepreneurial perspective, what experiences or lessons do you think can be shared currently?

A: If one wants to start a company now, the first requirement is to differentiate from existing players by possessing unique, specialized skills, as there is currently some homogenization in the market.

Secondly, I suggest identifying your specific application scenario before launching a startup. You may face funding challenges because this track has been active for over two years, with many investors entering the field or heavily backing several key players. Combining your unique characteristics and application scenarios to articulate a coherent development roadmap is something that takes time.

Q: We have noticed that Zhiyuan Robotics has received significant attention in Shanghai. In terms of company management, what is the greatest pressure you face? What are the most critical issues or problems to solve? Under the spotlight, what is the most important aspect for your team?

A: The most important aspect is already reflected in our values: product-centricity and talent-first.

Q: From Zhiyuan's perspective, what does the cost-reduction roadmap for hardware and software look like?

A: First, regarding the issue of cost reduction, we view it rationally. Zhiyuan's application scenarios and market direction are primarily focused on the B-end (business-to-business). From the perspective of replacing human labor, a reasonable price point is sufficient. From our standpoint, in industrial computing scenarios, as we continue to advance, many hardware components can be mass-produced via molding, bringing manufacturing costs down to relatively low levels. Therefore, we are not overly concerned about whether customers will accept the costs.

Q: What new ideas, progress, and methods have emerged in real-world data acquisition?

A: Regarding data, there are two main aspects: open-ended data collection and data collection from real-world application scenarios.

Many clients are willing to open up their environments for us to collect data, whether at their operational sites or inside retail stores. We integrate specialized commands into general-purpose datasets. Additionally, through data augmentation dimensions, we combine generative AI and simulation technologies. For the same piece of data, we modify environmental factors, lighting conditions, and even the texture and morphology of manipulated objects to perform extensive work aimed at increasing diversity.

In summary, I believe data collection is a continuous process that will not stop. Currently, many clients hope we will assist them in collecting data.

For instance, an appliance manufacturer has asked us to help collect data on dishwashers (loading dishes) and washing machines (washing clothes). We are receiving an increasing number of such specialized requests. We can aggregate this data with clients, turning it into one of our most valuable data assets.

Q: What are the advantages of traditional industrial robots when performing specific tasks for a particular product?

A: We are certainly not aiming to replace industrial automation. Those systems serve as specialized mechanisms that have been optimized to extreme performance levels for specific tasks. Our goal is to solve problems that traditional industrial automation cannot address.

Q: This year, thousands of robots are expected to be deployed. Which industries will they primarily flow into?

A: There are several, including industrial scenarios, scientific research and education scenarios, exhibition halls and retail stores, and other interactive scenarios.

Q: What is the rationale behind Zhiyuan's recent entry into the quadruped domain with its dog robot?

A: For quadruped robots, they fall under the Lingxi product line, which will also deploy some applications. Quadrupeds are currently a relatively mature product with lower costs and good stability. We have received many demands from the market side; they can serve as personal companions, conduct patrols overseas, guard homes, and so on, making them a very promising application scenario.

Why did we enter the market? Well, some competitors entered earlier and have already built up a significant revenue scale.

As for the issue of involution, we are not engaging in it for its own sake. Rather, through a greater variety of robot products and industry applications, we gain a deeper understanding of market demands for intelligence during the process. This also allows us to acquire more data and feedback, providing guidance for the development of humanoid robots.

Q: You previously mentioned that the industry is overheating. Will the quadruped market rely more quickly on consolidation?

A: The quadruped industry faces more intense competition and has a relatively lower barrier to entry.

Q: Will Zhiyuan enter special industries in the future? For example, patrol.

A: There are currently no such plans for special industries.

Q: I’d like to ask about Embodied AI. Does Zhiyuan have any plans to go global?

A: Yes. We already have some presence in North America, Europe, the Middle East, Japan and South Korea, and Southeast Asia. Many of these involve cooperation with local partners to achieve localization as part of our strategy for globalization.

At the same time, we began delivering our products overseas starting this year. As has been seen in many industries, China is a highly competitive market; products that succeed there can often make rapid breakthroughs abroad. This is indeed a key area we value.

Q: How does commercialization and implementation overseas differ from domestic operations?

A: Markets in Europe and North America tend to be more conservative than China’s. They seem less focused on new things, such as interactive scenarios, and instead prioritize applications in areas like industrial sectors to assess whether they meet the conditions for cost reduction.

Labor costs overseas are relatively high, and personnel management is challenging. Ensuring consistency and stability in factory products poses a significant challenge. In contrast, Chinese employees tend to be more stable and of higher quality.

Q: What are the attitudes of China and the US towards Embodied AI in the future?

A: Both China and the US place great importance on humanoid robots and Embodied AI. Many institutions analyze this as the final battle for hegemony between China and the US; whoever wins this battle will enter a new ecosystem.

The US has accumulated advantages in large models, technology, and talent to compete with us. However, this industry has distinct characteristics: it is highly dependent on hardware. This may be the challenge the US currently faces—the lack of a complete industrial chain capable of rapidly mass-producing high-quality robots. Compared to China, its manufacturing sector is not as developed, resulting in fewer application scenarios. I believe these are some of China's advantages.

Q: Regarding industrial capital, whether we invest in others or receive investments, what substantive progress has been made in cooperation with these industrial partners?

A: Yes, we often invest heavily in our upstream supply chains, such as sensors, joints, etc. Many of the investors and shareholders we have introduced are actually scenario providers, including automotive and 3C electronics sectors.

In fact, our scenario providers have already established close coordination around the upstream and downstream industries. We are integrating components from upstream supply chain enterprises into our complete machines, while downstream scenarios are opening up for us. Co-built projects have the potential to achieve delivery in the second half of the year.

Q: In robot development, which components might face supply issues?

A: Sensors, joints, computing chips, and batteries—these are roughly the categories.

Tesla's integration of domestic supply chains generally serves as a catalyst for the industry. In the past, the new energy sector in our country developed its entire supply chain system through the establishment of Tesla's Shanghai Gigafactory and requirements for domestic substitution. This foundation enabled the subsequent emergence of numerous domestic new energy brands. Moving forward, Tesla's technical standards for supply chains and its production volumes will benefit our industry.

The chip sector remains a sensitive topic. Currently, we utilize mainstream NVIDIA chips while also observing the development of domestic high-compute chips advancing toward 100T and 1,000T capabilities, which are adapting to mainstream models and providing a relatively secure environment.