RoboScience Machine Science completes A2 round of hundreds of millions of yuan in financing, as industrial capital continues to bet on embodied intelligence

The capital screening criteria for embodied intelligence are shifting from "telling a compelling story" to "delivering on it."

English text is being prepared.

English text is being prepared.

AI Tech Review has learned that on November 2, RoboScience Machine Science completed its A2 round of financing.

It is understood that this round was jointly led by Amber Capital, Hongruida Investment, and Shanghai Semiconductor Industry Investment, with 360 Strategic Investment, Shuang Capital, Guolian Minsheng Capital, Jinding Capital, and several listed company platforms participating as follow-on investors, with the financing amount reaching several hundred million yuan.

This means that this embodied company, which only officially began operations in March 2025, has already completed five rounds of financing in succession.

According to AI Tech Review, the financing market in the embodied industry has not been optimistic recently. Over the past year, the financing narrative in this track has changed several times: at first, you could get money just by talking about world models; later, talking about technology alone was no longer enough—you had to talk about implementation and the ROI of entering factories; and later still, you needed the imaginative space of "embodied RSI" (recursive self-improvement).

Behind this story inflation is the fact that the foundations of most teams cannot support the escalation of the narrative, and every time the trend shifts, they have to assemble a new set of talking points. Capital has accordingly become pickier: the same BP that could raise money last year cannot even get a meeting this year. (Leiphone)

But the rhythm of RoboScience Machine Science is completely different. It has continued to attract capital against the trend and has therefore become a sample company worth studying in the embodied field.

Looking back at the timing of Machine Science's past five financing rounds, around almost every round there was verifiable product and technology implementation: the ICRA Best Paper, the simulation platform RoboMirage, the general embodied large model Visics, the cloud embodied intelligence service EaaS co-created with Tencent Cloud, the first wheeled humanoid general-purpose robot REX G1, and real project implementation.

In a sense, this means that financing in the embodied market should not be a story of technology following financing; rather, financing should follow the delivery of technology.

In China's embodied industry, several clear factions can already be seen from the academic backgrounds of the founders of leading companies: the Tsinghua faction, the Berkeley circle, the Stanford circle, and so on.

Among them, the most influential is the Stanford circle.

Wang He of Galaxy General, Lu Cewu of Qiongche Intelligence, Su Hao of Hillbot, and () of RoboScience Machine Science—these companies and founders have different paths, but tracing upward, most of them can be traced to the same laboratory: the group at Stanford of Leonidas J. Guibas, member of the U.S. National Academy of Sciences, National Academy of Engineering, and American Academy of Arts and Sciences and geometric computing scholar. On the same lineage there are also OpenAI researcher Yan Mengyuan and Tsinghua University assistant professor Yi Li.

This is no coincidence.

Stanford is almost one of the birthplaces of global embodied intelligence research: Fei-Fei Li, as the initiator of ImageNet and a founder of the vision- and data-driven paradigm, Chelsea Finn's REAL Lab, and geometric computing scholar Leonidas J. Guibas, as well as founding members of the leading foreign embodied company Pi, all come from Stanford.

Around 2016, the geometric computing group of Professor Leonidas J. Guibas (member of the U.S. National Academy of Sciences, National Academy of Engineering, and American Academy of Arts and Sciences) at Stanford gathered a group of Chinese students, including Wang He, Shao Lin, and their senior fellow student Su Hao, among others.

Professor Guibas's group has a fairly consistent main line of research: starting from 3D vision, focusing on manipulation generalization, and believing in abstract representations and simulation. Shao Lin's personal trajectory is almost a complete sample of this main line.

Shao Lin entered Stanford in 2014. He first did 3D vision in the group of Leonidas J. Guibas, a titan of geometric computing and member of the U.S. National Academy of Sciences, National Academy of Engineering, and American Academy of Arts and Sciences, and later turned to robotic manipulation, studying under Jeannette Bohg. He is currently an assistant professor at the National University of Singapore.

Academically, Shao Lin's first-author paper UniGrasp is a benchmark work in data-driven dexterous hand grasping; in 2025, his team won the ICRA Best Paper Award in the robot manipulation and locomotion direction with D(R,O) Grasp from more than 4,000 submissions, becoming the first team from an Asian institution in nearly five years to win this award as the first affiliation, breaking the long-term monopoly of Europe and the United States; in 2026, Bi-Adapt was again shortlisted for the award, making it the only team in Asia to stand in this position for two consecutive years. Competing in the same field were Berkeley, MIT, Stanford, and Tsinghua.

In addition, the Stanford lineage has two other branches, namely the Fei-Fei Li line and the Andrew Ng line. The Fei-Fei Li line produced Lu Cewu and Wang Shiquan of Flexiv, and the Andrew Ng line produced Tian Ye.

In 2014, Tian Ye entered the Stanford AI Lab as the top student in the physics department of USTC, studying under Stanford AI godfather Andrew Ng, entering school in the same year as Shao Lin. He then spent seven years at Apple, rising to technical lead of the Apple AI Platform and becoming one of the youngest principal engineers at Apple headquarters, leading the creation of the AI platform known in the industry as "Apple's PyTorch and CUDA" as well as the construction of the world's first device-cloud collaborative large-model inference system; the latter was meant to serve not test machines in a laboratory, but more than 1 billion users and 2 billion devices, including iPhone, Vision Pro, and Apple Intelligence.

The core judgment this experience left him with is: Apple's moat has never been any single-point technology, but rather the system-level capability of integrated software and hardware.

Shao Lin and Tian Ye are old acquaintances; the photo is an early group shot.

At the end of 2024, one came out of Apple to run engineering, the other came out of academia to run research, and the two branches converged into RoboScience. Viewed within the Stanford lineage, this may be the startup with the most complete "academic depth + engineering depth" configuration: Shao Lin's line from UniGrasp to D(R,O) Grasp and then to Bi-Adapt corresponds to the model's cross-embodiment generalization capability; Tian Ye's experience at Apple doing device-cloud collaboration and serving billions of devices corresponds to turning capabilities into a deployable, iterable system. The combination of the two is another way of writing RoboScience's technology roadmap.

02

The embodied closed loop of three major milestones

From the release of the RoboMirage simulation engine in September 2025, to the release of the Visics general-purpose embodied large model in June 2026, and then to the release of the REX G1 wheeled humanoid robot in August, RoboScience spent about a year independently developing and connecting all three links—"data, model, and embodiment." This kind of technology combination is not unique to RoboScience, but among all companies doing full-stack work, AI Technology Review believes that RoboScience's technology choices have the clearest intent.

At the data layer, the first problem to solve is being able to "afford to eat." The biggest bottleneck in embodied intelligence is not algorithms, but data and environments; real-machine teleoperation collection is both expensive and slow, and the industry is generally trapped in the cycle of "data scale cannot support the Scaling Law."

RoboMirage started from the underlying solver and completed coverage of the entire object space in the embodied world, with efficiency and concurrency also reaching the industry's advanced level. Coupled rigid and flexible materials, the contact forces of extremely thin objects, and even actions that are difficult to complete via teleoperation, such as "shuffling" and "pulling a rubber band," can all be simulated with high precision. Combined with a fully automated video annotation pipeline, this "simulation + video" system pushes the cost of a single data item down to 0.05 yuan, about 1/20 to 1/200 of traditional solutions, and data is expanding at a rate of hundreds of thousands of hours per week. From early 2026 to October, video training data grew from 1 million to 4 million. The industry consensus now is that when a task has a sufficiently efficient and comprehensive validation environment, that task can often be solved by AI. RoboScience's data engine provides sufficient momentum for model iteration.

At the model layer, the core is the choice of "what to learn." Visics adopts the VLOA (Vision-Language-Object-Action) architecture, with dual internal engines: the upper-level world model is pretrained on massive internet video and is responsible for understanding tasks and previewing the 3D motion trajectories of objects; the lower-level manipulation model is responsible for translating trajectories into control signals for a specific robot. The interface in between is the object's 3D point-cloud trajectory.

The essence of this design can be summed up in one question: to teach a robot to grasp a cup, do you let it remember "how the cup should move," or "how this robot's joints should rotate"? The former is the universal language of the physical world; the latter is the dialect of specific hardware. A model that learns the former naturally has cross-embodiment capability, which is precisely the industrialization of Shao Lin's research line from UniGrasp to D(R,O) Grasp: one model drives LeapHand, Shadow Hand, and various other mainstream dexterous hands with different forms.

Whether the capability can withstand testing depends on real tasks.

In May 2025, RoboScience, based on the Visics general-purpose embodied large model, completed an extremely complex, high-precision, many-step furniture assembly task, all in one continuous shot. A few months earlier, on-site at WAIC 2026, AI Technology Review observed that RoboScience's booth demonstrated multiple dexterous hands of different brands running the same Visics general-purpose embodied large model, recognizing new hardware in 30 seconds with no fine-tuning, truly achieving plug-and-play hand swapping; for objects randomly handed over by visitors, the model autonomously grasped them throughout, achieving true generalization.

These demonstrations have an element of showmanship, but the two settings of "random objects" and "on-site hand swapping" are indeed a public test of generalization capability.

Therefore, at this conference, RoboScience and Sudo were the two companies that left the deepest impression on us.

At the embodiment layer, the final link of the closed loop. In August this year, RoboScience released the REX G1, which integrates mobility, perception, and dual-arm manipulation.

The significance of self-developed embodiment for RoboScience is not selling one more piece of hardware, but giving the flywheel of "simulation trains models, models drive embodiments, real data flows back" a carrier to run on, forming a true data flywheel.

Another external variable worth noting is the rapid iteration of general-purpose large models. The penetration of large models into the physical world is faster than most people expected. Models like GPT Astra can already understand spatial relationships and plan task steps, and some teams in the industry have begun directly using general-purpose large models as robots' "brains."

But if you have watched any livestreams or demos of large models controlling robotic arms, you will notice that the success rate is very low—once it lands in real-world scenarios, the shortcomings are quickly exposed. Ask a robot to move the cup on the table to the left, and a large model can break the task down beautifully: find the cup, plan the path, move the robotic arm. But is the cup glass or paper, how much force should be used, and what if it slips? Contact, friction, elasticity—these physical properties are not in text and video data; the model has never seen them, and it does not know them.

This is precisely the hardest part of embodied intelligence today: not "thinking it through," but "getting it done." Generalization of manipulation, execution speed, and task success rate determine whether robots can step out of demo videos. These require AI to learn the various physical properties of the real world and the corresponding data, and require companies to collect and build the relevant capabilities.

RoboScience's choice is to build this layer of capability solidly on its own.

Its Visics large model and RoboMirage simulation engine use a hierarchically decoupled architecture, combined with high-precision, high-diversity physical data, to train a unified manipulation model that spans embodiments, objects, and tasks. The robot first "understands" how objects move, and then, through the same interface, drives different bodies to complete tasks, no matter what those bodies look like.

03

Let's talk about RoboScience's commercialization approach

In a sense, we believe RoboScience's commercialization follows a dual software-hardware track.

The REX G1 wheeled humanoid robot delivers projects for specific scenarios, and at the same time, because the Visics general embodied large model adopts the VLOA technical route and a unique hierarchically decoupled design, Visics has become the industry's first cloud-based embodied large model. RoboScience has partnered with Tencent Cloud to jointly launch the first cloud-based embodied intelligence service, EaaS (Embodied-AI-as-a-Service), allowing users to obtain manipulation capabilities just like calling an API.

The two paths, one heavy and one light, together validate one thing: whether model capabilities can be reused, rather than developing every project from scratch.

At the same time, in designing its business model, RoboScience has laid out a diversified commercial landscape from cloud APIs to complete-machine delivery. It is not hard to see that this commercialization approach covers different groups.

The layer closest to the model is the cloud API.

Aimed at embodiment manufacturers, integrators, and developers, Visics is packaged into an extremely simple interface, available on demand, similar to cloud service providers, charging by API calls or licenses. For a company making dexterous hands or humanoid robots, what it has always lacked most is not hardware, but a brain that makes the hardware "able to work." Now it does not need to train models itself—calling an API gives it cross-object manipulation capabilities. What it is doing recently with Tencent Cloud's EaaS is precisely the lightest and easiest-to-scale layer.

The next layer down is installing a "brain" on third-party hardware—that is, providing third-party hardware with a plug-and-play "general manipulation brain."

Aimed at integrators in scenarios such as logistics, retail, hotels, general industrial automation, and equipment vendors, RoboScience provides a controller that integrates the model, computing power, and control interfaces, with a focus on plug-and-play. It targets the hard bones that traditional automation cannot crack: unordered sorting and mixed palletizing, where cases are piled together at random, of varying sizes and materials. Traditional solutions require modeling and customization one by one, whereas Visics, based on object trajectories, naturally does not care "what is being grasped." This layer charges by controller plus license and supports private deployment, effectively extending the capability boundary of automation equipment vendors: non-standard orders they could not take before, they can take now.

The heaviest layer is going in itself, targeting end scenarios such as instant retail front warehouses, supermarkets, and warehouse sorting.

RoboScience uses its self-developed REX G1 for complete-machine delivery, with perception, battery life, mobility, and manipulation all designed for scenario requirements, paired with a general end-effector interface, selling complete machines plus deployment and operations.

Why keep the heaviest layer for itself?

In our view, the answer lies in the data value of real-world scenarios. Retail and logistics scenarios have a vast variety of SKUs and ever-changing environments; they are both an exam room for testing the model's generalization ability and an entry point for accumulating real physical interaction data.

Equipped with the Visics general embodied large model, the REX G1 plays the role of closing this loop. Visics uses the 3D point cloud trajectories of objects as an intermediate representation, reducing the manipulation capability's dependence on specific goods and robot hardware, so that when the model faces unseen SKUs, it also has cross-embodiment, cross-object manipulation experience. It is understood that RoboScience is advancing multiple POCs with clients including large domestic retail and supply chain enterprises. For it, these projects are both commercialization validation and data entry points for continuous model iteration: successful experiences and failure cases in real operations can flow back into the training and evaluation system, driving model capability improvement.

From this perspective, investment in self-developed embodiments is expected to translate into compound interest for the entire technology system. The deeper the scenario validation, the fuller the data accumulation; the stronger the model capability, the higher the delivery value of APIs and controllers. With the three layers put together, the logic becomes clear: complete-machine delivery hones the model and accumulates data in the most complex scenarios; controllers bring validated capabilities into thousands of industries; and APIs turn it into a shared capability that anyone can call. Does that sound a bit like an embodied version of OpenAI?

04

Three rounds of financing in one year, industrial capital continues to bet

Finally, let us remind everyone to pay attention to the composition of investors in this round of RoboScience. Looking through the financing structure of this round, we can find that in addition to financial institutions, Amber Capital, Shuang Capital, the 360 Strategic Investment Fund, and several listed company platforms also appear—the presence of industrial capital is very notable.

Industrial capital is closer to the application front lines, and its questions are more specific: how long can the equipment run, does it need someone to watch over it, and how many days does it take to debug a new category of goods. Their entry can be understood as the industrial side betting on the path of "reusable capabilities," but investment does not equal orders; ultimately, the site still has to speak for itself. Judging from the results, RoboScience has already completed the due diligence test of industrial capital.

Back to the financing itself: the story inflation in the embodied intelligence industry is essentially because most teams lack chips that can be sustainably cashed in. RoboScience's consecutive financing shows that it has built a complete chain among research, engineering, and products. After the A2 round, the evaluation criteria will be more plain: whether the robot can work stably after entering the site, how much extra it costs to switch to another item or another piece of equipment, and whether delivery becomes lighter as projects increase.

In the second half of 2026, all the new financing stories that embodied intelligence can tell have been told. Next comes a new stage that requires hardcore technology and real deployment scenarios to speak. (Leiphone)