After the Embodied AI Sector Cools Down, This Stanford Startup Secures Hundreds of Millions in Funding Against the Trend

Author: Wu Qiong, Investment Circle

At a turning point for embodied intelligence, a financing deal emerges against the trend—today (November 2), RoboScience announced the completion of a Series A2 funding round of hundreds of millions, led by Amber Capital, Hongruida Investment, and Shanghai Semiconductor Industry Investment, with participation from Shuang Capital, 360 Venture Capital, Guolian Tongbao, Jinding Capital, and several listed companies.

Looking into RoboScience, it is backed by two classmates from Stanford University’s AI Lab—Tian Ye and Shao Lin. During their time on campus, they often discussed how robots could enter human life. Years later, witnessing the explosive growth of embodied intelligence in China, they chose to return to their homeland to start a business, leading to the birth of RoboScience.

Stanford is almost one of the global sources of embodied intelligence research: Li Fei-Fei, the initiator of ImageNet, the pioneer of vision and data-driven paradigms, Chelsea Finn’s REAL Lab, geometric computing scholar Leonidas J. Guibas, and the founding team of leading foreign embodied companies Pi all hail from Stanford. Around 2016, Professor Leonidas J. Guibas (an academician of the American Academy of Arts and Sciences) gathered a group of Chinese students in his geometric computing group, including Wang He, Shao Lin, and their junior Su Hao.

From 3D geometry and robot learning to physical interaction, Stanford’s research accumulation has provided important inspiration for these entrepreneurs exploring general embodied intelligence. This is also reflected in the technical genes of RoboScience: currently, the company’s self-developed Visics general embodied large model has broken through the ontology generalization problem; the self-developed physical simulation engine RoboMirage further overcomes data thresholds. On the path to general embodiment, RoboScience has chosen to directly target the most challenging points.

At this moment, investors' choices often carry significant directional meaning. Seizing the opportunity of completing the financing, the investment community engaged in an in-depth conversation with RoboScience’s founder and CEO Tian Ye, slowly unfolding a different embodied startup story from most others.

Two Stanford Classmates Join Forces to Tackle the Generalization Problem in Embodiment

The Stanford connection is the first impression that RoboScience leaves on most people.

The team’s inception began at Stanford—Tian Ye graduated from the University of Science and Technology of China and went to Stanford for further studies in 2014, studying under Professor Andrew Ng, thus entering the world of AI.

This period of study became one of the most valuable experiences in Tian Ye’s life. At that time, deep learning was just emerging globally, and Andrew Ng was one of its earliest advocates. Under his guidance, Tian Ye not only encountered the world’s cutting-edge technologies but also engaged in numerous attempts to apply these technologies to real-world scenarios.

Many great entrepreneurial stories in the business world often begin with the purest friendships formed on campus—also in this geeky and innovative environment, Tian Ye and his classmate Shao Lin forged a bond. The difference was that Shao Lin studied under embodied intelligence expert Jeannette Bohg and geometric computing master Leonidas Guibas, focusing on the field of robot learning.

Although their research directions were different, their curiosity unexpectedly resonated: how to enable AI capabilities to enter the physical world and interact with robots using general natural language? "Since our time at Stanford, we have been discussing when robots can truly enter human mass production and life," Tian Ye recalled to the investment community.

However, at that time, the concept of embodied intelligence had not yet exploded, and the timing for entrepreneurship was not mature. Therefore, Tian Ye chose to first join Apple, playing a core role in building Apple’s AI product ecosystem and AI application ecosystem, growing from an engineer to a technical leader, leading the world’s first end-cloud collaborative large model inference system Apple Intelligence and the core platform CoreML and CreateML, known as "Apple’s PyTorch and CUDA"; Shao Lin joined the National University of Singapore’s Computer Science Department as an assistant professor, continuing research in robot learning and dexterous manipulation.

These two distinctly different experiences completed the final piece of the entrepreneurial puzzle: starting from Stanford, Tian Ye accumulated engineering experience in deploying technology to large-scale devices and users at Apple; the series of research led by Shao Lin also became the technical foundation for their future entrepreneurship. Thus, a complementary entrepreneurial combination of "academia and engineering" was formed.

Photo of the two founders of RoboScience (Shao Lin on the left, Tian Ye on the right)

Fast forward to 2024, the generality of large language models exploded with the emergence of ChatGPT, and both academia and industry began to attempt to replicate this Scaling Law in the physical world of robots. Finally, the opportunity arrived, and Tian Ye decisively left his job to return to China to start RoboScience. The two continued to fight side by side—Tian Ye as CEO and Shao Lin as Chief Scientist.

In the meantime, the domestic embodied intelligence sector completely exploded, with various capital and top entrepreneurs entering the fray. Looking back at this wave, the Stanford connection stands out—Hillbot founder Su Hao, Galaxy General founder Wang He, Qiongche Intelligent founder Lu Cewu... they all come from the Stanford robotics and AI academic lineage. As Tian Ye feels, "In the AI field, Stanford is undoubtedly at the forefront globally, with countless academic and engineering innovations originating from here." Quietly, a "Stanford Legion" has emerged in China’s embodied intelligence sector.

The higher the starting point, the greater the vision. RoboScience has chosen a more challenging path: through a full-stack technology route that integrates software and hardware, developing both a general embodied large model and the robot body as well as physical simulation.

"The most important things in embodied intelligence are two: models and data," as Tian Ye said, this basically explains the growth path of RoboScience— to break through the generalization problem faced by embodied intelligence, the team developed the Visics general embodied large model.

Specifically, the Visics general embodied large model adopts the VLOA (Vision-Language-Object-Action) architecture, using Object Trajectory (3D point cloud trajectory of objects) as a unified intermediate representation standard. This design completely decouples "cognition" from "execution"—the upper-level embodied world model is responsible for predicting and deducing the physical motion trajectory of objects, while the lower-level general operation model converts the trajectory into control instructions adaptable to any robot body.

At the same time, the company developed a high-precision simulation engine, RoboMirage. Former Tesla AI director and OpenAI co-founder Andrej Karpathy proposed in his article "Verifiability" that when a task has a sufficiently efficient and comprehensive verification environment that allows for repeated trial and error, efficient trial and error, and can automatically provide reward feedback, the likelihood of AI solving the problem through continuous training will greatly increase.

RoboMirage has achieved high concurrency, real-time performance, and coverage of the entire object space. This provides significant momentum for the development of RoboScience’s embodied models. Combined with fully automated video annotation and cleaning pipelines, it has built a "simulation + video" dual data flywheel. Tian Ye shared a set of data: RoboScience’s data route has reduced the cost of acquiring a single piece of data to 0.05 yuan, down to 1/20 to 1/200 of traditional solutions. From early 2026 to October, RoboScience’s video training data increased from 1 million to 4 million and continues to expand, fundamentally breaking through the dual bottleneck of data scale and cost, creating conditions for the scaling law of embodiment.

As general large language models continue to demonstrate spatial understanding capabilities, the capital market’s inquiries into embodied intelligence companies have become increasingly direct: as the capability boundaries of general models continue to expand, what is the independent value and long-term barrier of embodied companies developing their own "brains"?

Tian Ye’s judgment is that general large language models excel at tool invocation, task decomposition, and intent orchestration, but from understanding tasks to driving robots to stably complete operations, there remains a physical world threshold. For embodied companies, the core value lies in whether they can enable different hardware to complete operations in complex environments with sufficient speed and high success rates, and reuse this capability across more bodies and tasks. Behind this, it requires models to learn physical laws such as contact, friction, and elasticity, as well as continuously accumulate physical interaction data that is difficult to obtain directly from text and video, allowing the system to efficiently and robustly complete real tasks in the physical world.

This is also the underlying logic behind RoboScience’s layout of the Visics large model and RoboMirage simulation engine: to scale the generation of physical interaction data through the simulation engine, reducing the data cost of model training; to connect object motion understanding and robot operation through a layered decoupling architecture, reducing the dependence of model capabilities on specific hardware. Together, they point to a commercially valuable goal: to enable the same set of operational capabilities to be reused across bodies and tasks, reducing the marginal costs of adapting new hardware and deploying new scenarios. From a venture capital perspective, the imaginative space of this route lies in transforming the continuous accumulation of data and models into scalable, deliverable general operational capabilities, and serving more robot manufacturers and application developers through a unified Visics model interface.

After achieving significant leaps on both the model and data sides, RoboScience released its first generation of robot body—the REX G1 wheeled humanoid general robot, which can be widely applied in logistics, retail, factories, homes, and other scenarios.

Like Apple, the real barrier has never been a single point technology, but a systemic capability that integrates software and hardware. RoboScience is no different—by building full-stack capabilities, it aims to create the next generation of robots capable of autonomously executing complex tasks at a faster iteration speed.

Counter-Trend Financing: One Year to Complete the Full-Stack Loop of Embodiment

Investors had noticed this pair of Stanford classmates early on.

In the AI era, investors often establish connections with high-level executives from large companies and university professors who have entrepreneurial potential, so they can invest as soon as these individuals decide to start a business. This scene also happened with Tian Ye—he recalls that even before he and Shao Lin left their jobs, investors began to approach them, maintaining regular communication.

Everything unfolded naturally. As soon as RoboScience was established, a seed round of several tens of millions of RMB, exclusively invested by Zero One Ventures, was quickly secured. Two months after completing the financing, Shao Lin’s team won the Best Paper Award in the field of robot manipulation and motion with their general dexterous grasping framework D(R,O) Grasp, becoming the first team from an Asian institution to receive this honor as the first author in nearly five years. This research also serves as the technical source for RoboScience’s breakthrough in cross-body dexterous manipulation problems.

Investors have continued to increase their stakes: in July of the previous year, nearly 200 million yuan was raised in an angel round led by JD.com, with participation from China Merchants Venture Capital and SenseTime's Guoxiang Capital, while existing shareholder Zero One Venture Capital also increased their investment; in February 2026, several hundred million yuan was raised in a Pre-A round, with investors including PwC Capital and Dacheng Wisdom, among other well-known investment institutions, while China Merchants Venture Capital and Zero One Venture Capital continued to increase their stakes; up to now, the company has completed multiple rounds of financing totaling several hundred million yuan in Series A funding.

RoboScience has maintained a steady and pragmatic pace in its financing efforts. In fact, after experiencing a boom in financing in the first half of the year, the embodied intelligence sector has noticeably cooled in the second half of this year, with investors becoming increasingly cautious and investment standards tightening. The lower the industry goes, the more significant the financing becomes.

What has convinced investors to commit is the milestones achieved by RoboScience at each financing stage: in September 2025, the high-precision simulation engine RoboMirage was launched, enabling a low-cost, large-scale data flywheel; in June 2026, the Visics general-purpose embodied large model was officially launched, demonstrating cross-ontology, cross-object, and cross-task generalization capabilities with the VLOA architecture; just two months later, the first wheeled humanoid robot, REX G1, made its debut. In other words, within a year, key links such as data, models, and robot bodies were connected, forming a comprehensive embodied full-stack closed loop.

It is worth mentioning that before completing this round of financing, RoboScience also reached a strategic cooperation with Tencent Cloud to create the world's first cloud-based EaaS (Embodied-AI-as-a-Service) embodied intelligence service. Based on this cooperation, Visics became the world's first cloud-based embodied large model, open for public access and utilization.

Behind this lies RoboScience's contemplation of commercialization. Tian Ye revealed that the team currently has three industrialization routes: providing overall embodied solutions for end applications; offering model services EaaS that can be called from the cloud, as well as controllers deployed on the edge, opening model capabilities to a broader range of developers, ontology companies, and scene integrators.

In Tian Ye's words, these three routes balance "depth and breadth": the full-stack route delves into key scenarios, while the controllers and EaaS extend capabilities, thus covering a wide range of application scenarios that a single company cannot fully grasp.

These factors have also impressed investors. As Amber Capital stated: "The iteration of embodied intelligence technology is accelerating, transitioning from technical validation to large-scale application across different fields, with the general capabilities of open scenarios and landing costs being key to industrial breakthroughs. RoboScience has built a full-stack technical system that connects data, models, and ontology around this core proposition, further enhancing cross-ontology, cross-object, and cross-task generalization capabilities, achieving continuous data accumulation and capability iteration in real working conditions. We highly recognize the team's forward-looking judgment on the technical path and systematic execution capabilities, and we are optimistic about the company's accelerated validation in real scenarios, promoting the large-scale reuse of model capabilities, gradually growing into an important infrastructure platform for the era of general embodied intelligence."

In the face of the challenges of commercial mass production and large-scale validation, the actions of industrial capital are often seen as a barometer. For all players, such investors are invaluable—not only for financial support but also for providing real validation scenarios.

Looking at the shareholder lineup of RoboScience, such faces are not uncommon: JD.com has dozens of "Asia No. 1" smart logistics parks and thousands of warehouses across the country; behind Shuang Capital is the Hong Kong billionaire Li Jinji family, which has a vast production base and supply chain network globally. Their businesses are filled with numerous non-standard, complex, and difficult-to-automate processes, which are precisely ideal replacement scenarios for embodied intelligence.

This round of financing has also seen multiple industrial capital and listed company platforms banding together to cast trust votes for RoboScience. Tian Ye revealed that internal cooperation is underway with a group of industry players, which will be announced in succession soon.

This is undoubtedly the best testing ground for embodied companies. Moving forward, the acceleration path for RoboScience is becoming increasingly clear: the funds from this round of financing will primarily support the continuous iteration and upgrading of large models, advancing POC validation and large-scale landing in core scenarios such as logistics, retail, and industry. At the same time, it will further strengthen the reserve of high-end talent.

Towards Generalized Endgame

Achieving general embodied intelligence is almost the ultimate dream of all embodied entrepreneurs.

This is an extremely large and romantic proposition. But generalization means that robots can perceive the environment, make autonomous decisions, and learn from feedback after acting, just like humans, and there are still many engineering challenges to overcome behind this, with a long road ahead. When imagination and reality diverge, the industry begins to shift from "universal fantasy" to "specialized precedence". Some even argue that "generalization is still too far away; deeply cultivating vertical scenarios is the key to survival."

However, in Tian Ye's view, this is not a binary choice. "The core of general embodied intelligence lies in reducing marginal costs, which is also our core competitiveness in the future." He further explained that the reason RoboScience insists on developing full-stack products is to achieve deep optimization in specific fields such as logistics and retail.

"We definitely need to land in certain scenarios first, but what supports it is an evolving, increasingly generalized, and decreasingly costly underlying technology." Tian Ye emphasized that a positive sustainable business closed loop needs to be deeply coupled with scenarios, while long-term competitiveness requires expanding scale by continuously reducing marginal costs.

First, we will enter real high-value, high-demand industries, and then let robots move towards broader scenarios, ultimately achieving the general embodied intelligence that everyone expects. On this road, generalization is the hardest nut to crack.

In simple terms, most current embodied models can achieve nearly 100% task success rates after sufficient data collection and specialized training in fixed scenarios. However, as soon as the operating objects change or the environment is slightly altered, the task success rate of the robots can drop significantly.

In response to this challenge, RoboScience has provided its own solution: leveraging its self-developed VLOA structure, it now supports generalization across three dimensions: cross-ontology, cross-object, and cross-task. A memorable moment was at the 2026 World Artificial Intelligence Conference, where RoboScience demonstrated "30-second rapid handover"—staff directly replaced different brands of dexterous hand hardware without needing to re-adapt or fine-tune, with the dexterous hand being instantly recognized within 30 seconds and smoothly executing various grasping tasks, truly achieving "ready to use upon handover," validating its cross-ontology generalization capabilities.

Additionally, the dexterous hand equipped with the Visics general embodied large model successfully grasped any items provided by the audience on-site; even more groundbreaking was that at the WRC held in August this year, RoboScience simultaneously opened the reasoning process of the world model at the booth, becoming the first company in the industry to publicly demonstrate real-time visual reasoning of the world model.

"Embodied intelligence is highly related to the physical world, requiring the synchronization of real scenarios, real tasks, and algorithms to evolve together." Tian Ye is not blindly optimistic, stating, "It is definitely a step-by-step process, from small to large, gradually moving from specific industries to broader sectors." But he is also confident that this will happen in the coming years.

As the wave of embodied intelligence entrepreneurship continues, controversies and bubbles have emerged. When discussing this, Tian Ye appears quite calm, stating, "Bubbles may arise from everyone trying different approaches, much like biological evolution, where some disappear during the process while others survive. This is a natural phenomenon of technological evolution."

In the second half of the embodied era, companies that solely rely on flashy demos to tell stories will face increasing survival pressure. Tian Ye prefers to focus his energy on "getting things done," stating, "Short-term fluctuations are just noise for us. Returning to the essence of business, what everyone will focus on next is how technology can truly land in products."

The road to AGI in the physical world is still long. Looking back years from now, this moment may well be a watershed.

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.