The archive · AI & Models · Strategic decision · 2025–
OriginFlow bets muscle signals, not video, can feed robots the data they lack
Tsinghua PhD Qin Shentao founded OriginFlow in Aug 2025; three rounds raised ¥500M+ to turn human muscle signals into robot training data
OriginFlow（渊澈太初）
What the business is
OriginFlow builds embodied-AI data infrastructure: non-invasive sEMG wristbands plus a foundation model (PULSE) that turns human muscle signals, vision and motion into 'Human Tokens' usable for training robots, starting with industrial and home-service scenarios.
Starting capital：Over ¥500 million across angel, strategic and Pre-A1 rounds in about five months (Dec 2025–May 2026), backed by Lanchi Ventures, Oasis Capital, Monolith, 58 and others
How it started
Qin Shentao, born 2001 in rural Shanxi, studied mechanical engineering at Harbin Institute of Technology, where his robotics teams won national competitions and he became the first junior to win both the President's Medal and the May 4th Medal. He incubated at Li Zexiang's Shenzhen X-Innovation and Kai-Fu Lee's MiraclePlus before starting a Tsinghua PhD, and founded OriginFlow in August 2025 after concluding that model architecture would converge — the real moat would be data infrastructure.
What happened
OriginFlow completed angel (co-led by Lanchi Ventures and Oasis Capital), strategic (58, Puhua Capital, Tsinghua alumni seed funds) and Pre-A1 (led by Monolith) rounds within roughly five months, raising over ¥500 million with more than ten investors; Lanchi reportedly sent a term sheet the same day as the first meeting. It shipped the OriginKitGen 1.0 wristband, showed PULSE 0.2 reconstructing fingertip force in real time, and started data projects with manufacturing partners and 58 Group for home-robot skill databases.
No ending yet — it is still running.
Background
OriginFlow (渊澈太初) is a Beijing-based embodied-AI company founded in August 2025 by Qin Shentao, a 25-year-old Tsinghua PhD student. Its bet is that robots are starved of physical-interaction data — text built large language models, fleets built self-driving cars, but no one has an equivalent source for touch, force and grip — and that the gap can be filled by reading the muscle signals behind human movement.
The core technology, called NeuroScale, is a non-invasive motor-neural interface: a wristband captures surface electromyography (sEMG) signals, and the company's PULSE foundation model reconstructs hand pose, contact force and tendon drive into 'Human Tokens' that robots can learn from. By sampling nerve intent rather than just watching action results, it aims to capture what vision-only teleoperation misses — exactly when a hand tightens, how much force it applies, how it adjusts to a wobbly object.
Qin, who won national robotics competitions at Harbin Institute of Technology and later incubated with Li Zexiang and Kai-Fu Lee's programs, decided during his Tsinghua PhD that model architecture would converge and data infrastructure would decide the winner. OriginFlow raised over ¥500 million across three rounds (angel, strategic and Pre-A1) in about five months, backed by Lanchi Ventures, Oasis Capital, Monolith, 58 and others; one investor says it sent a term sheet the same day as the first meeting.
The company is pursuing industrial manufacturing and home services, partnering with 58 Group to collect real household operation data for home-robot skill libraries. Its stated arc is three phases: From Human (collecting and encoding human movement), With Human (always-on AI hardware), and Enhance Human — a data company today, a potential human-machine interface platform tomorrow.
What has to be true
- Data is the bottleneck: LLMs had the internet and AVs had fleets, but embodied AI has no equivalent corpus of force and touch — OriginFlow targets that exact gap.
- Signal advantage: sEMG captures motor intent before the action, reconstructing force and grip that vision-only pipelines miss, with non-invasive, low-cost hardware.
- Scale path: ordinary people wearing wristbands in daily life can generate training data far cheaper than robot teleoperation or sim-to-real transfer.
- Proven founder signal: a national robotics champion and Tsinghua researcher whose pedigree made investors race — Lanchi issued a term sheet the same day.
- Multiple shots: the same neural-interface stack could become a consumer human-machine interface, not just a data tool.
What can be applied
In a data-starved model era the scarce asset is the data pipeline: OriginFlow industrialized a signal nobody had, and capital chased the wedge before the product was finished.
Aftermath
As of mid-June 2026, OriginFlow had raised over ¥500 million in three rounds within five months and was building its NeuroScale pipeline: the OriginKitGen 1.0 wristband, the PULSE model (0.2 demonstrated continuous hand tracking and fingertip-force reconstruction; 0.3 in development), and data infrastructure with ORACLE auto-annotation and CHORD alignment. It had started industrial data projects with manufacturing partners and, with 58 Group, was building skill databases for home robots, aiming to scale human-action data toward trillions of hours and eventually ship consumer wearables.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card