RoboScience Machine Science completes Series A2 funding of hundreds of millions of yuan; Stanford classmates start a business in embodied intelligence

At a turning point for embodied intelligence, a financing round has emerged against the trend—

Today (November 2), RoboScience announced the completion of a Series A2 financing round of several hundred million yuan, jointly led by Amber Capital, Hongruida Investment, and Shanghai Semiconductor Industry Investment, with Suang Capital, 360 Strategic Investment, Guolian Tongbao, Jinding Capital, and several listed company platforms participating.

Looking into RoboScience, behind it are two classmates from the Stanford University AI Lab—Tian Ye and Shao Lin. As early as their campus days, they often discussed together how robots could enter human life. Until years later, witnessing China's embodied intelligence reach an explosive inflection point, they chose to return to China to start a business, and RoboScience was born.

Stanford is almost one of the birthplaces of global embodied intelligence research: as the initiator of ImageNet, Fei-Fei Li, the founder of the vision- and data-driven paradigm, Chelsea Finn's REAL Lab, geometric computing scholar Leonidas J. Guibas, and the founding team of leading overseas embodied company Pi all come from Stanford. Around 2016, the geometric computing group of Stanford Professor Leonidas J. Guibas (member of all three U.S. national academies) gathered a group of Chinese students, including Wang He, Shao Lin, and their junior fellow student Su Hao.

From 3D geometry and robot learning to physical interaction, Stanford's research accumulation also provided important inspiration for this group of entrepreneurs exploring general embodied intelligence. This is also reflected in RoboScience's technological DNA: the company's self-developed Visics general embodied large model breaks through the challenge of embodiment generalization; its self-developed physics simulation engine RoboMirage further breaks through the data threshold. On the path to general embodiment, RoboScience has chosen to aim directly at the hardest barrier.

At this moment, investors' choices often carry more of a bellwether significance. Taking advantage of the financing milestone, Investment World conducted an in-depth conversation with Tian Ye, founder and CEO of RoboScience, and an embodied entrepreneurship story different from most previous ones gradually unfolded.

Two Stanford classmates join forces to start a business

Relentlessly tackling the hard problem of embodied generalization

Stanford lineage—this is the first impression RoboScience leaves on most people.

And the team's germination began precisely at Stanford—Tian Ye graduated from the University of Science and Technology of China as an undergraduate and went to Stanford in 2014 for further study, studying under Professor Andrew Ng, and thus entered the world of AI.

This period of study became one of the most valuable experiences of Tian Ye's life. At that time, global deep learning was just emerging, and Andrew Ng was one of its earliest advocates. Under his teacher's guidance, Tian Ye not only came into contact with the world's most cutting-edge technology but also made numerous attempts to apply technology to real-world scenarios.

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

Although their research directions differed, the two unexpectedly shared the same frequency of curiosity: how to bring AI capabilities into the physical world, using general natural language to interact with robots and perform operations? "Since Stanford, we have been talking about when robots can truly enter large-scale human production and life," Tian Ye recalled to Investment World.

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

It was precisely these two very different experiences that completed the last piece of the entrepreneurial puzzle: starting from Stanford together, Tian Ye then accumulated at Apple the engineering experience of deploying technology to large-scale devices and users; a series of studies led by Shao Lin's team also became the technological foundation for future entrepreneurship. Thus, an entrepreneurial combination complementary in "academia and engineering" was formed.

*Photo of the two founders of RoboScience (left: Shao Lin; right: Tian Ye)

In the blink of an eye, by 2024, the generality of large language models exploded with the emergence of ChatGPT, and both academia and industry began trying to replicate this Scaling Law onto robots in the physical world. Finally, the opportunity arrived, and Tian Ye decisively resigned and returned to China to start a business, founding RoboScience. The two continued to fight side by side—Tian Ye as CEO and Shao Lin as chief scientist.

Around this time, the domestic embodied intelligence track exploded completely, with capital from all sides and top entrepreneurs entering the field. Looking back at this wave, the Stanford lineage is particularly striking—Hillbot founder Su Hao, Galaxy General founder Wang He, Qiongche Intelligence founder Lu Cewu... They all come from Stanford's robotics and AI academic lineage. As Tian Ye feels, "Looking across the AI field, Stanford is unquestionably a global frontier, and countless academic and engineering innovations have happened here." Quietly, a "Stanford legion" has risen in China's embodied intelligence track.

The higher the starting point, the greater the vision. RoboScience Machine Science chose a harder path: through a full-stack technology route integrating software and hardware, it not only develops its own general embodied large model, but also builds the robot body and physical simulation.

"The two most important things in embodied intelligence are models and data," as Tian Ye said. This basically explains RoboScience Machine Science's growth path—in order to break through the generalization challenges facing embodied intelligence, the team developed its own 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 trajectories of objects) as the unified intermediate representation standard. This design completely decouples "cognition" from "execution"—the upper-layer embodied world model is responsible for predicting and simulating the physical motion trajectories of objects, while the lower-layer general manipulation model converts trajectories into control commands adapted to any robot body.

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

RoboMirage has already achieved high concurrency, real-time performance, and coverage of the entire object space. This provides extremely high momentum for the development of RoboScience's embodied models. Combined with a fully automated video annotation and cleaning pipeline, it has built a "simulation + video" dual data flywheel. Tian Ye shared a set of figures: RoboScience Machine Science's data route has pushed the acquisition cost of a single data item down to 0.05 yuan, reduced to 1/20 to 1/200 of traditional solutions. From the beginning of 2026 to October, RoboScience Machine Science's video training data grew from 1 million to 4 million and continues to expand, fundamentally breaking through the dual bottlenecks of data scale and cost and creating conditions for the Scaling Law of embodied intelligence.

As general large language models continue to demonstrate spatial understanding capabilities, the questions capital markets ask embodied intelligence companies are becoming more direct: as the capability boundaries of general models continue to expand, where exactly lies the independent value and long-term moat of embodied companies developing their own "brains"?

Tian Ye's judgment is that general large language models are good at tool calling, task decomposition, and intent orchestration, but from understanding a task to driving a robot to stably complete an operation, there is still a threshold in the physical world. For embodied companies, the core value lies in whether they can enable different hardware to complete operations in complex environments at sufficiently fast speeds and with sufficiently high success rates, and reuse this capability across more bodies and tasks. Behind this, models need to learn physical laws such as contact, friction, and elasticity, and also need to continuously accumulate physical interaction data that is difficult to obtain directly from text and video, so that the system can efficiently and robustly complete real tasks in the physical world.

This is also the underlying logic of RoboScience Machine Science's layout of the Visics large model and the RoboMirage simulation engine: use the simulation engine to generate physical interaction data at scale and reduce the data cost of model training; use a hierarchically decoupled architecture to connect object motion understanding with robot manipulation and reduce the model's dependence on specific hardware. Both point together to a commercially valuable goal: enabling the same set of manipulation capabilities to be reused across bodies and tasks, and reducing the marginal cost of adapting to new hardware and deploying in new scenarios. From a venture capital perspective, the imaginative space of this route lies in converting the continuous accumulation of data and models into general manipulation capabilities that can be delivered at scale, and serving more robot manufacturers and application developers through a unified Visics model interface.

After huge leaps on both the model side and the data side, RoboScience Machine Science released its first-generation robot body—the REX G1 wheeled humanoid general-purpose robot, which can be widely used in logistics, retail, factories, homes, and other scenarios.

Like Apple, the real moat is never any single-point technology, but systematic capability integrating software and hardware. The same is true of RoboScience Machine Science—by building full-stack capabilities and iterating faster, it aims to create next-generation robots capable of autonomously performing complex tasks.

In the AI era, investors often establish contact in advance with senior executives at major companies and university professors who have entrepreneurial potential, so that they can invest at the first opportunity when these people decide to start businesses. This scene also happened with Tian Ye—in memory, before he and Shao Lin had even left their jobs, investors were already coming to them one after another to maintain regular exchanges.

Everything fell into place naturally. Just as RoboScience Machine Science was founded, a seed round of tens of millions of yuan, exclusively invested by Zero One Venture Capital, was quickly secured. Two months after completing the financing, Shao Lin's team won the Best Paper Award in robot manipulation and motion with the general dexterous grasping framework D(R,O) Grasp, becoming the first team from an Asian institution in nearly five years to receive this honor as the first affiliation. This research is also the technical origin of RoboScience Machine Science's breakthrough in cross-body dexterous manipulation.

Since then, investors have continued to increase their commitments: in July of that year, the company completed a nearly 200 million yuan angel round led by JD.com, with China Merchants Venture and SenseTime Guoxiang Capital participating, and existing shareholder Zero One Venture Capital increasing its investment; in February 2026, it completed a Pre-A round of several hundred million yuan, with investors including well-known institutions such as Puhua Capital and Fortune Capital, while China Merchants Venture and Zero One Venture Capital continued to increase their commitments; and now, the company has completed multiple A-series rounds totaling several hundred million yuan.

RoboScience has always maintained a steady and pragmatic pace in financing. In fact, after the financing boom in the first half of the year, the embodied intelligence sector clearly cooled in the second half of this year, with investors becoming more cautious and investment standards stricter. The deeper the industry trough, the more weight financing carries.

What made investors firm in their commitment was the milestones RoboScience achieved at each financing node: in September 2025, it released the high-precision simulation engine RoboMirage, enabling a low-cost, large-scale data flywheel; in June 2026, it officially launched the Visics general embodied large model, using the VLOA architecture to prove generalization across embodiments, objects, and tasks; and just two months later, its first wheeled humanoid robot, REX G1, was unveiled. In other words, within one year it connected key links such as data, models, and robot bodies, forming a full-stack closed loop for general embodied intelligence.

It is worth mentioning that before completing this round of financing, RoboScience also reached a strategic cooperation with Tencent Cloud to build the world's first cloud-based EaaS (Embodied-AI-as-a-Service) embodied intelligence service. On the basis of this cooperation, Visics became the world's first cloud-based embodied large model, opening its calling ecosystem to the public.

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

In Tian Ye's words, these three routes balance "depth and breadth": the full-stack route deeply cultivates key scenarios, while controllers and EaaS extend capabilities outward, thereby covering vast application scenarios that a single company cannot fully master.

All of this is also what moved investors. As Amber Capital stated: "Embodied intelligence technology is iterating faster and is going through the stage from technical validation to large-scale application in different fields. General capabilities in open scenarios and deployment costs are key to industrial breakthroughs. Centering on this core proposition, RoboScience has built a full-stack technology system connecting data, models, and bodies, further enhancing generalization across embodiments, objects, and tasks, and 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 capability, and are optimistic that the company will accelerate validation in real scenarios, promote large-scale reuse of model capabilities, and gradually grow into an important infrastructure platform in the era of general embodied intelligence."

Facing the test of commercial mass production and large-scale validation, moves by industrial capital are often regarded as a bellwether. For all players, such investors are extremely valuable—not only for financial support, but also for providing real validation scenarios.

Looking at RoboScience's shareholder lineup, 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 Lee Kum Kee family, a Hong Kong billionaire dynasty, with a huge global production base and supply chain network. Their businesses are full of large numbers of non-standard, cumbersome, and hard-to-automate processes, which are precisely ideal scenarios for embodied intelligence to replace.

This round of financing also saw multiple industrial capital and listed company platforms form a group to cast votes of confidence in RoboScience. Tian Ye revealed that the company is internally launching cooperation with a group of industrial parties, which will be announced successively soon.

This is undoubtedly the best testing ground for an embodied company. Next, RoboScience's acceleration path is becoming clearer: the funds from this round will focus on supporting continuous iteration and upgrading of the large model, advancing POC validation and large-scale deployment in core scenarios such as logistics, retail, and industry. At the same time, it will further strengthen its reserve of high-end talent.

Moving toward the general endgame

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

This is an extremely vast and romantic proposition. But generality means robots can perceive the environment, make decisions autonomously, and learn from feedback after acting, learning while working like humans. Behind this, there are still many engineering challenges to overcome, and there is still a long way to go. When imagination and reality diverge, the industry begins to move from the "omnipotent fantasy" to the "dedicated first" stage. Some even argue that "generality is still too far away, and deepening vertical scenarios is the key to survival."

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

"We will certainly land in certain scenarios first, but what supports it is the underlying technology that is constantly evolving, becoming more general, and increasingly lower in cost." Tian Ye emphasized that a positive and sustainable commercial closed loop requires deep coupling with scenarios, while long-term competitiveness requires expanding scale by continuously reducing marginal costs.

Start with real high-value, high-demand industries, then let robots move into broader scenarios, and ultimately reach the general embodied intelligence everyone expects. On this path, generalization is the hardest nut to crack.

Simply put, most current embodied models can achieve a task success rate close to 100% in fixed scenarios after sufficient data collection and specialized training. But as soon as the object being manipulated is changed, or the environment is slightly altered, the robot's task success rate drops sharply.

Facing this challenge, RoboScience has offered its own solution: with its self-developed VLOA architecture, it already supports generalization across three major dimensions: cross-embodiment, cross-object, and cross-task. One particularly memorable moment was at the 2026 World Artificial Intelligence Conference, where RoboScience staged a live "30-second rapid hand swap"—staff directly replaced the dexterous hand hardware with different brands, with no need for re-adaptation or fine-tuning. The dexterous hand was recognized instantly within 30 seconds and smoothly performed multiple types of grasping tasks, truly achieving "swap the hand and it works," validating cross-embodiment generalization capability.

In addition, the dexterous hand equipped with the Visics general embodied large model was able to successfully grasp any object provided by audience members on site; even more groundbreaking, at this year's WRC held in August, RoboScience simultaneously opened up the world model's reasoning process at its booth, becoming the industry's first company to publicly demonstrate real-time visualized reasoning of a world model.

"Embodied intelligence is highly related to the physical world, and it needs to evolve in sync with real scenarios, real tasks, and algorithms." Tian Ye is not blindly optimistic. "It will definitely be step by step, from small to large, gradually realized from specific industries to broader industries." But he is also certain that this will happen in the next few years.

Amid the embodied intelligence startup wave, controversy and bubbles have followed. Speaking of this, Tian Ye instead seemed very calm. "Bubbles may come from everyone making different attempts, just like biological evolution—some disappear during the evolutionary process, and some survive. This is a natural phenomenon of technological evolution."

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

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

This article is from the WeChat public account "Investment World" (ID: pedaily2012), author: Wu Qiong, published by 36Kr with authorization.