Building infrastructure for Physical AI is a marathon.
Recently, ScanArk Technology (Shanghai) Co., Ltd. (hereinafter referred to as "ScanArk") announced the completion of its Pre-Angel round of financing.
With this, the company has completed three rounds of financing—Seed, Seed+, and Pre-Angel—within four months, raising a cumulative total of tens of millions of CNY.
Founded in November 2024 and registered in Yangpu, Shanghai, ScanArk is a Reality-to-Sim embodied intelligence infrastructure company focused on Physical AI, dedicated to building spatial data infrastructure that connects the real world with artificial intelligence.
The core technical team of ScanArk comes from the Robotics and Artificial Intelligence Laboratory (RAIL) at Tongji University.
The origins of this laboratory can be traced back to 1990. As one of the earliest domestic teams engaged in robotics research, it has long focused on autonomous mobile robots, humanoid robots, 3D vision, and multi-sensor fusion. The lab has undertaken scientific research tasks such as the National High Technology Research and Development Program of China (863 Program), the National Science and Technology Support Program, and key projects of the National Natural Science Foundation of China. Its related achievements have won honors including the National Science and Technology Progress Award, the First Prize of the Shanghai Science and Technology Progress Award, and the First Prize of the Shanghai Natural Science Award.
In short, behind this young entrepreneurial team lies more than 30 years of accumulated research in robotics.
Peng Yun, the Founder and CEO, was born in 1998. His technical experience spans industrial 3D vision, 3D spatial understanding, and simulation data production. He previously led core projects in the 3D Grid Algorithm Team at Huawei's 2012 Laboratories and has published multiple papers on 3D vision as the first author. In 2020 and 2021, he won the championships in both the Standard Platform League and Domestic League at the RoboCup China Open with the TJArk team.
Shuai Su, Co-founder and CTO, is responsible for hardware architecture and algorithm R&D, focusing on SLAM and 3D reconstruction. He participated in the early R&D of NIO's multi-sensor fusion calibration algorithms and conducted research on visual SLAM based on lightweight deep learning at OPPO XR. He has published multiple papers as the first author at conferences such as ECCV Workshops, ICRA, and IROS.
Chuangwei Liu, Co-founder and Co-CTO, leads the R&D of 3DGS reconstruction algorithms and generative models, focusing on 3D vision and binocular depth perception. He was the first to introduce visual foundation models into the field of binocular stereo matching. His paper ViTAStereo, as the first author, reported achieving first place in the KITTI Stereo 2012 benchmark.
The three are not only PhDs from the same laboratory and fellow students under the same supervisor, but also teammates who debugged robots and refined algorithms side by side on the RoboCup competition field.
This long-term collaboration, spanning from academia to competitions and then to entrepreneurship, means their technical tacit understanding was not forged through temporary integration, but grew out of repeated joint debugging sessions.
This experience is also preserved in the company's name. The "Ark" in ScanArk is inherited from the TJArk team name—it not only means "ark" but also echoes Tongji University's motto of "rowing in the same boat."
Creating a World That Robots Can Enter
What ScanArk is doing can be summarized in one sentence: transforming the real world into a digital twin that robots can enter, act in, and learn from.
This sounds like an upgraded version of simulation modeling, but there is a fundamental difference between the two. Traditional simulation environments typically select tables, chairs, cabinets, and home appliances from an asset library and arrange them according to rules. Such a room looks complete and reasonable, but it loses the characteristics of a real living environment—the logic of object placement, occlusion relationships, and signs of wear and tear.
These messy, exception-filled environments are exactly what robots will ultimately have to face.
To address this, ScanArk has pioneered a brand-new 3DGS (3D Gaussian Splatting) format. Through self-developed models, it generates real-world environments into native 3DGS scenes with physical and interactive information. While restoring visual and spatial structures, it generates object attributes, part relationships, and interaction states.
Geometric scale, layout, navigability, lighting, reflections, clutter, and usage traces are all preserved. Only by grounding in reality can the experience gained in simulation have a better chance of transferring to real-world deployment.
Building on this, real-world observations are organized into visual, geometric, semantic, physical, and scene relationships, supporting the simulation of collisions, contacts, joint movements, and state changes. This is then used for navigation, grasping, manipulation, and policy evaluation, and finally validated through real-robot execution, failure diagnosis, and real-world calibration.
In other words, what customers receive is not a set of viewable 3D assets or a one-time dataset, but an environment where robots can repeatedly enter, attempt, fail, and start over. Engineering teams can reset initial states, alter task conditions, compare the outcomes of different actions, investigate the causes of failure, and benchmark simulation performance against real-robot results.
As new real-world observations and deployment failures continuously feed back, the digital world is continuously updated. ScanArk aims to form a closed loop from reality to simulation and back to reality, making the digital twin a space for robots to accumulate experience.
Centered on this technical roadmap, the company is building real-digital environment infrastructure for robot interactive learning and world model training, providing embodied intelligence teams with generative, editable, and interactive training environments and evaluation foundations.
Physical AI Moving Towards Reality Still Needs to Fill the Data Infrastructure Gap
The track ScanArk has entered becomes clearer when viewed within a broader industrial context.
As AI evolves from digital content processing to perception, decision-making, and action in real-world environments, Physical AI is receiving increasing attention.
Many consider 2026 to be the inaugural year of Physical AI. However, it has yet to experience its own "ChatGPT moment."
The core gap preventing the current leap in Physical AI capabilities lies not in model scale or hardware performance, but in the absence of an infrastructure capable of continuously generating, accumulating, and feeding back high-quality physical interaction data.
According to estimates from brokerage research reports, training an embodied intelligence model with practical application capabilities requires the industry to have at least 10 million hours of multimodal interaction data, whereas the current globally accumulated data volume accounts for less than 5% of the demand.
The explosion of ChatGPT in late 2022 relied on decades of accumulated text data on the internet. Physical AI does not have such conditions. The data it needs is not just about "seeing," but also about acting in real-world environments, receiving feedback, and adjusting strategies. There is a gap that is difficult to cross directly: massive amounts of video can be used to pre-train visual capabilities, but they cannot directly output the control commands required for robots to execute tasks.
This requires forming a positive cycle: acquiring real-world data through robot deployment, using the data to train foundation models, and then expanding the deployment scope with improved model capabilities, ultimately forming a flywheel of "deploying more robots, acquiring more data, training stronger models, and entering more scenarios."
Before this flywheel starts spinning, Physical AI will remain confined to customized projects and testing phases.
What ScanArk is doing fits precisely into this gap.
It holistically restores real living and working environments into interactive digital twins, allowing robots to repeatedly execute actions, observe outcomes, and accumulate experience within them. For embodied intelligence teams, this means they can conduct large-scale interactive learning training in the digital world without having to trial and error everything in the real environment.
Based on the pace of completing three rounds of financing in four months, the capital market's attention to this technical roadmap is genuine. However, building infrastructure for Physical AI is a marathon. Whoever can first achieve a breakthrough in data production efficiency and simulation transfer accuracy will hold the key entry point for Physical AI to move from the laboratory to industrialization. ScanArk has just begun its journey.
Text | Zhouyu Editor | Yida