ScanArk raises tens of millions of yuan across three rounds in four months
What's the deal? ScanArk (景立构)Dealroom has a profile for this one. Try Dealroom →, a Chinese startup building Reality-to-Sim infrastructure for embodied AI, has closed a Pre-angel round, its third raise in four months. The company has pulled in tens of millions of yuan in total across seed, seed+, and Pre-angel rounds. Investors include Guohai Innovation CapitalDealroom has a profile for this one. Try Dealroom →, Xiaomiao LangchengDealroom has a profile for this one. Try Dealroom →, a leading embodied-robotics company, and well-known industrial investors.
What's the endgame? ScanArk turns real-world environments into executable, interactive digital copies where robots can act, fail, and learn. Using in-house models, it generates native 3D Gaussian Splatting (3DGS) scenes that carry physical and interaction data alongside visual and spatial structure.
Why it matters: Embodied AI faces a data bottleneck. Internet images and first-person video capture rich visuals but lack spatial structure, physical state, and action feedback; real-robot collection is accurate but costly and hard to scale across long-tail environments.
ScanArk's pitch is to close that gap. Its environments preserve real layouts, object relationships, and wear, and respond to actions through simulated collision, contact, and joint movement — then calibrate against real-robot testing.
Why now? The robot-simulation sector is heating up as embodied models grow more capable, raising demand for digital environments that both mirror the real world and react to actions. The company bets that interaction learning — where robots gather experience by acting, observing results, and updating strategy — becomes the next phase for embodied AI.
The field: ScanArk competes across converging approaches. Rivals include reconstruction-focused players, World Labs on the generative side, and Lightwheel, which builds SimReady assets and has disclosed work training GeelyDealroom has a profile for this one. Try Dealroom → humanoid robots. Nvidia, meanwhile, links real-scene reconstruction into Isaac Sim via NuRec and extends environments with Cosmos.
The signal: As foundation models shift from passive pretraining to learning through feedback, the race is on to build the "digital training grounds" where robots practise. ScanArk's core team, drawn from Tongji University's Robotics and Artificial Intelligence Lab, is wagering that interactive, visually faithful simulation is the infrastructure that unlocks it.
Read more: 36kr.com
Image credit: Generated with Gemini
