Building physical reality in the world of tensors.
Author Zanzhu
Handcrafted Quantity丨100%
Founded in July this year, TensorWorld Physical Technology has consecutively completed angel round and angel+ round financing within a month, raising over 100 million dollars in total.
In the angel round, it was jointly led by Kaihui Fund and Sequoia China, with follow-on investments from IDG Capital, Dinghui VGC, Xinglian Capital, and Ant Group; the angel+ round was led by Ant Group, Heyu Capital, and Meituan's strategic investment, with continued follow-on investments from Dacheng Caizhi, NIO Capital, Jinghe Capital, and multiple old shareholders.
How could a newly established company rapidly gain the collective backing of top capital with two consecutive rounds of financing?
The answer may not only lie in the strategic direction of Physical AI but also in its choice of a more fundamental technological path: starting from the laws of physics, constructing a physically native base model, allowing AI to truly understand the physical world.
This choice is backed not only by the company’s forward-looking judgment of the industry but also by a team that has long been engaged in physical modeling, geometric representation, embodiment, and large model engineering.
Forward-looking Industry Judgment
From non-consensus to the current focal issue
The Physical AI industry has undergone a tremendous transformation this year, from the world model that surged in popularity at the beginning of the year, to the rise of World Action Model, and then to the release of Astra by OpenAI in September, signaling that the trends in the embodied and physical intelligence industry seem to be in perpetual flux. How to find one’s anchor amidst the noise of the market and technology is a significant test of the founder's judgment, which is also the most critical ability for a startup.
Under increasingly impressive demos, there has been another underlying thread gradually bringing the foundational technology of physical simulation to the forefront: In June, Fei-Fei Li publicly stated that the essence of the world is geometry, physics, and dynamics, and simulation is a bridge connecting vision and action. Though the public pays the least attention to it, it is the most influential; in July, World Labs acquired the robotic simulation company SceniX, further confirming the positioning of this world model-centric enterprise in the simulation field; in October, SuperMimic revolutionized the view on dynamic simulation in the embodied field, demonstrating that good neural simulators can achieve hundreds of times the speed of intelligent emergence.
This series of developments released a common signal: the competition in Physical AI is gradually returning to underlying capabilities. The scaling up of large-scale embodied data and the construction of high-performance and high-precision physical base models are defining the slope of the new era's scaling curve. The physical representation, simulation, and dynamics pretraining that were seldom highlighted in the past are becoming the focal issues in the industry.
Furthermore, the judgment of the founding team at TensorWorld came long before physical simulation became the industry focus. In 2023, founder Wu Haixu first proposed the concept of Physics-Token in Transolver, which has been followed for a long time by NVIDIA and many overseas startups, making the physical modality an important research direction independent of text and vision; subsequent works like Transolver++ and Transolver-3 series further achieved industrial-grade physical simulation accuracy; earlier this year, Wu Haixu collaborated with MIT’s Professor He Kaiming to propose GeoPT, using synthetic dynamics to construct pre-training signals that enable models to autonomously learn transferable physical laws, making the scaling of physical foundational models possible.
Looking back at this research system, TensorWorld aims to answer not just a Policy Model or some demos but to constantly pursue the core expressions of the physical world. From representation definition to backbone network design and large-scale pre-training, underlying innovation is the technical essence of this company. As stated on the company’s official website, “Building physical reality in the world of tensors.” is the company's original intention.
Full-stack Team
Defining physical reality for the first time in the AI community
If industry foresight allows a company to lead by half a step, a full-stack team and talent density is the core reason for maintaining the company’s acceleration. The industry also values TensorWorld for its exceptionally rare “complementary” star founding team.
Founder and CEO Wu Haixu, as the first author of GeoPT, successfully introduced Scaling Law into physical foundational models, allowing the industry to see the possibility of inherent physical laws in foundational model “emergence.” Born in 1998, Wu Haixu entered Tsinghua University’s School of Software in 2016, subsequently pursued a second degree in pure and applied mathematics in the Mathematics Department, and conducted research at Harvard University in 2019. After graduating, he remained at Tsinghua to continue his doctoral studies under Associate Professor Long Mingsheng. In 2025, he will pursue postdoctoral research at MIT CSAIL, with a co-advising mentor being pioneer Wojciech Matusik in the field of computational design.
Doctor Wu Haixu has just graduated, with over 30,000 citations on Google Scholar, and has received the special award from Tsinghua, being included in the top 2% of global scientists list. His research spans dynamic modeling, geometric representation, and physical foundational models, making pioneering contributions in multiple fields. Early on, his Autoformer and TimesNet focused on dynamic process modeling, becoming essential readings in AI time series analysis; subsequently, research shifted toward physical spatial representation, proposing Physics-Token in 2023 and advancing this field to the era of physical foundational models with GeoPT this year.
His research trajectory has gradually propelled the development of physical foundational models in the past few years, addressing key questions such as how to represent physical information, how to learn physical laws through pretraining, and whether model capabilities can continuously expand with data and computation scale.
If Wu Haixu provides the core capability for TensorWorld to build physical foundational models, the other co-founders bring strengths from engineering, embodied intelligence, and geometric modeling, enabling TensorWorld to tap into a broader market for Physical AI.
Zheng Wen, Ph.D. graduate from Stanford’s Computer Science Department, studied under computational physics pioneer Stanley Osher. Before joining TensorWorld, he served as Vice President of AI Technology at Kuaishou and head of the AI laboratory Y-tech. His experience covers cutting-edge AI research, product engineering for hundreds of millions of users, and management of computing infrastructure at the scale of hundreds of thousands of cards, capable of transforming research outcomes into industrial-grade products and system capabilities.
Ye Weirui, Ph.D. graduate from Tsinghua University’s Institute for Interdisciplinary Information Research, then conducted postdoctoral research at MIT and visited UC Berkeley’s Pieter Abbeel lab. He focuses on generative world models and embodied intelligence, with his EfficientZero, proposed in 2021, being a classic work in Model-based RL and selected for Sequoia China · MIT Technology Review AI25 (pioneers in AI innovation under 25 years old) in 2026.
Zhou Junsheng, a doctoral student at Tsinghua University’s School of Software, mainly engages in 3D spatial intelligence research and was selected for the Baidu Scholarship. His participation in Uni3D built a billion-level 3D representation, selected for ICLR 2024 Spotlight, providing important support for the team in three-dimensional space understanding and geometric representation.
Meanwhile, the team’s academic research experience connects leading universities such as Tsinghua, MIT, Stanford, and UC Berkeley, covering directions like physical simulation, computer vision, machine learning, and embodied intelligence. They have long-term academic collaborations with world-class scholars such as Wojciech Matusik, He Kaiming, Michael Jordan, and Pieter Abbeel from MIT and UC Berkeley.
The four co-founders each have backgrounds enough to “open a table on their own,” yet now they gather at TensorWorld, collectively tackling the most fundamental issues in Physical AI.
Focused on embodiment, engineering, and scientific exploration
Building a shared physical intelligence base
Returning to the focus of TensorWorld on "physical reality," why is this issue so important?
For a long time, people's attention has always been drawn to the performance of agents, the swinging of a robotic arm, and the generation of a novel vehicle model are far more interesting than the environments it interacts with and the actual airflow it encounters. However, what truly determines whether embodiment can be smoothly deployed and whether industrial design can go into production is still the feedback from the physical world. What TensorWorld focuses on is how to transform the rules of the physical world into intrinsic capabilities of foundational models and based on such capabilities, construct physically native foundational models to achieve more accurate understanding, more precise simulation, and more stable operations.
The so-called "physical reality" means "no difference between prediction and the real world." What TensorWorld hopes to build is not just a video generation model or game engine dressed in the guise of a world model but rather a physically native simulator capable of continuously approaching the real world. "Reality" is not "visually reasonable," but rather when geometric structures, material properties, boundary conditions, or external forces change, the state changes deduced by the model can still match the real world without difference.
Once the paradigm of TensorWorld is successfully established, the physical foundational models they construct will become infrastructure shared among embodiment, engineering, and scientific exploration.
For example, robots can scale up practice for tasks that are difficult to collect, costly, or even dangerous before actual operation; engineers can simulate the performance of different designs under stress, flow, heat transfer, and deformation conditions before manufacturing physical prototypes; scientists can propose hypotheses, simulate experiments, and discover new laws from the discrepancies between predictions and empirical evidence within larger parameter spaces. Physical AI will no longer be limited by the speed of real machine data collection and the modeling cost of traditional simulation. As recently demonstrated by SuperMimic, the iteration and emergence of physical intelligence will be accelerated hundreds, even thousands of times.
Physical reality is the ultimate criterion
For a model that only acts in the digital world, visual reasonableness is already a significant advancement; but for a model that aims to enter the physical world and serve embodiment, industry, and research, appearing correct is ironically the most dangerous "false prosperity." What truly matters is whether it can accurately predict dynamics in the real world and whether it can still successfully deduce results under unfamiliar geometric structures, materials, and stress conditions.
As physical reality increasingly becomes the industry focus, the long-term value of the route chosen by TensorWorld at its founding in July this year is becoming more apparent.
When AI moves from the digital world to the physical world, the standards of intelligence are not determined by declarations, but provided by the real world.
Building physical reality in the world of tensors. This is TensorWorld’s answer.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到[email protected]，本平台相关工作人员将会进行核查。
