RoboScience’s Visics Model Is Teaching a Robot to See the World’s Physics

The robot hand, a five-fingered gripper mounted on a wheeled base, approaches a small cardboard box. It isn’t a box the system has been trained on, and the gripper isn’t the specific hardware the model was initially tuned for. Yet the hand pauses, adjusts its angle, and closes its fingers with a pressure calibrated to lift, not crush. This moment, a simple pick-and-place, is the entire thesis of RoboScience. It’s not about building a better single-purpose robot for a factory line. It’s about building a brain that can understand the physics of a box, the friction of cardboard, and the mechanics of a gripper it may have never physically touched before [RoboScience] [Pandaily, June 2026].

The wedge is cross-embodiment

RoboScience’s core bet is on what it calls “cross-embodiment general-purpose embodied intelligence” [Dealroom]. The goal is a software layer,their proprietary Visics model,that can be deployed across different robot bodies, or embodiments, to perform universal manipulation tasks. The technical architecture, as described by the company, is built on a dual-engine system. One engine, an Embodied World Model, is trained on massive internet video data to learn object physics, motion patterns, and force dynamics. The other, a General Operation Model, translates those learned trajectories into actions for a specific robot’s hardware [Pandaily, June 2026]. The commercial promise is to sell intelligence, not just iron. The company offers this as a cloud-based Embodied-AI-as-a-Service (EaaS) and licenses its models to robotics manufacturers and automation enterprises, while also developing full robot systems like its REX series of wheeled humanoids [RoboScience].

A team built for the layers

The founding team reads as a deliberate assembly of the three core competencies required to pull this off: frontier AI research, industrial hardware scaling, and capital mobilization.

• AI and platform pedigree. CEO Tian Ye holds a master’s from Stanford’s AI Lab and is described in investor materials as a former Apple AI-platform lead who deployed infrastructure across billions of devices [PERPLEXITY SONAR PRO BRIEF]. Chief Scientist Shao Lin, a Stanford classmate of Ye’s, is an assistant professor at the National University of Singapore, anchoring the academic research frontier [PERPLEXITY SONAR PRO BRIEF].
• Hardware and supply chain. Co-founder Liu Penghai brings over 20 years of hardware R&D and management, previously serving as a vice president at the Ecovacs Group and general manager of its motor division [RoboScience]. This is experience in taking complex consumer robotics from design to volume production.
• Financial architecture. Co-founder Wang Tao, formerly head of fundraising at SenseTime Guoxiang Capital, helped mobilize multi-billion-yuan funds [RoboScience]. His role speaks to the capital-intensive nature of the bet and the need to navigate China’s deep-tech investment landscape.

The $100M+ momentum

Since its founding in December 2024, RoboScience has moved with remarkable velocity, securing over $100 million across multiple rounds from a who’s who of Chinese tech and financial investors [PANews, November 2026]. The financing history illustrates a stacking of strategic partners.

Angel Round (Jul 2025) | 200 | M CNY
Pre-A Round (Feb 2026) | Undisclosed |
Series A (May 2026) | 1000 | M CNY
Series A2 (Nov 2026) | Undisclosed |

Key investors include JD.com, which led the angel round; SenseTime Guoxiang Capital, the investment arm of the AI giant; and a later-stage cohort featuring Shanghai Semiconductor Industry Investment and Amber Capital [PANews, November 2026] [LinkedIn / 01VC, August 2025]. The presence of JD.com and semiconductor investors hints at potential future commercial and hardware synergies, though the available evidence does not confirm specific deployment partnerships.

The long road to general purpose

The ambition is staggering, and the path is lined with formidable, well-documented challenges that extend beyond pure software innovation. The market context notes that successful adoption “will depend on pairing innovation in AI capabilities with equal rigor in engineering discipline, safety thinking, and governance” [SAE World Congress 2026]. Furthermore, industry observers caution that early 2026 deployments show embodied AI can be “a flexible tool to selected workflows, not proof that a general-purpose robot workforce has arrived” [iTechGuides, 2026]. For RoboScience, the primary risks are not about funding or team, but about the fundamental difficulty of the problem.

• The simulation-to-reality gap. The company’s differentiable multimodal physical simulation engine, RoboMirage, is crucial for training [RoboScience]. However, the leap from even the most advanced simulation to reliable, safe operation in the unpredictable physical world remains a monumental engineering hurdle.
• Economic proof. In sectors like healthcare, analyses point to “substantial acquisition and maintenance costs… combined with limited economic evidence” as significant barriers to adoption [Healthcare, March 2026]. RoboScience must prove its EaaS or robot solutions deliver unambiguous ROI in complex, real-world environments.
• The commercialization clock. The company has moved from founding to nine-figure funding in under two years. That capital brings pressure to transition from R&D to commercial deployments and named customer references, which are not yet visible in the public record.

The company’s answer to these challenges appears to be its layered approach. By focusing on a licensable AI model and a cloud service, it can potentially find revenue and validation points with hardware partners before its own full robots need to conquer the world. It is a bet on intelligence-as-a-layer first, robot-as-a-product second.

In the end, RoboScience is answering a question that has lingered since the first wave of robotics hype: what if the limiting factor was never the arm or the wheel, but the brain’s inability to understand the world it was built to manipulate? The company is betting that the next breakthrough isn’t a better actuator, but a model that has watched, learned, and internalized the physics of a million YouTube videos, ready to guide a hand,any hand,to pick up a box it has never seen before.

Sources

1. [RoboScience, retrieved 2026] RoboScience company website | https://www.robosci.com/
2. [Pandaily, June 2026] Article on Visics model architecture | https://pandaily.com/
3. [Dealroom, retrieved 2026] RoboScience company profile | https://dealroom.co/companies/roboscience1/
4. [PERPLEXITY SONAR PRO BRIEF, retrieved 2026] Company and team briefing |
5. [PANews, November 2026] After the Embodied AI Wave Cooled, This Stanford Startup Raised... | https://www.panewslab.com/en/articles/01a19a32-9a7a-776e-8ccb-166704357b90
6. [LinkedIn / 01VC, August 2025] RoboScience Raises RMB 200 Million in Angel Round | https://www.linkedin.com/posts/01vc_roboscience-raises-rmb-200-million-in-angel-activity-7356968488046915584-f9XZ
7. [SAE World Congress 2026, retrieved 2026] Commentary on embodied AI adoption |
8. [iTechGuides, 2026] Analysis on embodied AI deployments |
9. [Healthcare, March 2026] Analysis on barriers to embodied AI in healthcare |