Most robot AI models today are trained for one body and one set of cameras, which means a model that works on a wheeled platform often has to be retrained from scratch for a bipedal one. RoboScience, a Beijing-based embodied-AI startup founded in December 2024, is betting that gap is worth hundreds of millions of yuan to close. The company announced a Series A2 round on November 2, 2026, led by Amber Capital, Hongruida Investment and Shanghai Semiconductor Industrial Investment, according to TechStartups. Shuang Capital, the strategic investment arm of security-software company 360, Guolian Tongbao and Jinding Capital also joined.
RoboScience's core product, Visics, is a general-purpose manipulation model the company says can control different robotic systems without separate programming for each environment. The company pairs that with RoboMirage, a simulation platform for training and testing robot behavior before it reaches physical hardware, and REX G1, a wheeled humanoid-style robot that serves as RoboScience's own reference platform for testing the software stack.
Betting on portability over bespoke hardware
RoboScience's pitch sits inside a broader argument playing out across China's embodied-AI sector: that the hard problem left to solve is not building better robot bodies but building models flexible enough to transfer across the many bodies already on the market. Chinese embodied-AI companies raised roughly $13 billion across 322 deals in the first half of 2026 alone, five times the volume from the same period a year earlier, as hardware makers increasingly outsource their intelligence layer to specialist software firms rather than building it in-house.
That dynamic puts RoboScience in competition with a crowded field of Chinese foundation-model startups making similar cross-platform claims, as well as with vertically integrated players like Unitree and AgiBot that build both the hardware and the brain. RoboScience's answer is to lean on neutrality: by not committing to one hardware form factor, Visics can in principle sell into fleets built by companies RoboScience doesn't compete with directly, similar to how Physical Intelligence and Skild AI have positioned themselves in the US market.
The simulation-to-reality bet
RoboMirage addresses a problem that has slowed deployment across the industry: training robots on real hardware is slow and expensive, and mistakes made during training can damage equipment or disrupt the facilities where testing happens. A simulation layer lets RoboScience iterate faster on Visics before any code reaches a physical arm, though the approach only pays off if behaviors learned in simulation transfer reliably to messy, unpredictable real-world settings — the same sim-to-real gap that has challenged robotics labs for years.
RoboScience also struck a partnership with Tencent Cloud for embodied-AI services, giving it distribution and compute infrastructure beyond its own balance sheet. With China's embodied-AI funding pace cooling somewhat in the second half of 2026 after the H1 surge, RoboScience's ability to convert this round into commercial deployments across multiple robot makers, rather than just its own REX G1, will determine whether the model-agnostic strategy actually works at scale.