"The robotics industry has never lacked impressive demos; what it lacks are systems that can work reliably in the real world long-term." On August 22, these words from Tsingyan Technology CTO Wang Fan circulated via Economic Information Daily and Sina Finance, alongside the announcement of a nine-figure RMB Series A round.
Funding Overview
- Company: Tsingyan Technology (Beijing) Co., Ltd., jointly incubated by Tsinghua University and the Beijing Yanqi Lake Applied Mathematics Institute
- Round: Nine-figure RMB Series A
- Lead investor: Legend Capital (君联资本)
- Co-investors: Suzhou Venture Capital (苏创投), Ningbo Jilian Fund (宁波继联基金)
- Use of funds: Physical AI infrastructure and foundation model R&D
- Tsingyan follows a slower but deeper-moat path: world model infrastructure + physics engine + geometric representation, closer to infrastructure than to selling "brains" or data.
- It joins an August cluster of world-model/data-layer deals: Zhongke Diwuji (over 1B RMB), Moushen Intelligent (500M Pre-A), Tsingyan (nine-figure A), Mifeng (nine-figure Pre-A), Liulanshuzhi (nine-figure).
- Valuation pressure is real: Chinese embodied AI valuations currently price in shipment volumes and demo appeal, not infrastructure patience. But the 93.5B RMB half-year funding pool suggests room for contrarian routes.
The round is notable less for its size than for its timing and technical positioning. It lands in the shadow of Unitree's first-day gain of 460.34% on the STAR Market (August 19) and H1 2026 funding of 93.5 billion RMB in China's embodied intelligence sector (up 5x year-on-year, with deal count up 137%). While most embodied AI companies pitch "VLA models + large models + data flywheels," Tsingyan is betting on a "geometry-physics-driven physical AI framework."
Key Points from Wang Fan's WRC 2026 Talk
Wang broke the gap between demo and scaled deployment into three independent barriers, each mapped to a framework component:
1. Generalization → Geometric-physics representation. Most VLA models learn "what things look like" from data but degrade when objects, environments, or tasks change — Unitree founder Wang Xingxing acknowledged this at WRC. Tsingyan's approach builds 3D/4D world modeling on differential-geometric invariants (rotation, translation, scaling) and physical constraints (dynamics, conservation laws, contact), so the representation embeds hard constraints on how objects can move. The model never needs to learn from scratch that an apple falls down.
2. Reliability → Verifiable learning. Tsingyan claims its in-house physics engine achieves 5.2x the efficiency of mainstream physics engine Newton in dynamic distance and collision computation at equal precision, and an order-of-magnitude advantage in soft-body deformation. This makes "world model predicts, verifier checks" an engineering reality: models can self-validate predictions during training rather than waiting for post-deployment feedback — shifting reliability from post-hoc statistics to in-process calibration.
3. Continual learning → Continued evolution. Three capabilities: meta-learning (learning how to learn), long-term memory (cross-session state retention), and autonomous knowledge discovery. If it works, robots no longer need to "return to the factory for updates" — after six months in a plant, they master neighboring workstations on their own, a break from the "factory-fixed + periodic OTA" model.
Capital Market Context
Roadmap
Wang outlined three layers: infrastructure and model layer (physical AI infrastructure, models, data systems), verification layer (verifiable learning paradigm + efficient physics engine), and evolution layer (deepened continual learning).
The announcement marks a genuine route fork in the sector: "data + parameters" versus "geometry + verification + continual learning" now run in parallel. The latter has a higher ceiling and deeper moat, but a much slower cadence — making it the line worth tracking through the second half of 2026's physical AI route debate.