The archive · Space, Robots, Defence · Strategic decision · 2025–2026
Genesis AI bets hands-first robotics: $105M seed, GENE-26.5 full-stack model
Genesis AI bet dexterous manipulation is robotics' core problem and built a human-sized hand, data glove and foundation model together.
Genesis AI
What the business is
Genesis AI builds foundational AI for robots: the GENE foundation-model family trained across vision, tactile, proprioception and action data, plus the human-sized Genesis Hand, a wearable data-collection glove and simulation, aimed at general-purpose manipulation.
Starting capital：$105M seed co-led by Eclipse Ventures and Khosla Ventures (announced July 2025), with Bpifrance, HSG and individuals including Eric Schmidt, Xavier Niel, Daniela Rus and Vladlen Koltun.
How it started
Genesis AI was created in early 2025 by Zhou Xian (CEO) and Théophile Gervet (president, formerly a research scientist at Mistral AI). Just months after founding, in July 2025, it emerged from stealth with a $105M seed round co-led by Eclipse and Khosla Ventures — with Bpifrance, HSG, Eric Schmidt, Xavier Niel, Daniela Rus and Vladlen Koltun also in — to build a 'data engine' for training robot models.
What happened
The founders concluded a model alone could not reach human-level manipulation, so Genesis went full stack: it designed the human-sized Genesis Hand with partner Wuji Tech, added a lightweight sensor-loaded glove and egocentric video capture so daily jobs double as data collection, and built a simulation system to speed model iteration. In May 2026 it released GENE-26.5, its first foundation model, demoing cooking, lab pipetting, smoothie-making, wire harnessing and a Rubik's cube solve at 1x speed with one set of weights; the team grew to ~60 across Paris, California and London.
No ending yet — it is still running.
Background
Genesis AI was founded in early 2025 by Zhou Xian (CEO) and Théophile Gervet (president, formerly a research scientist at Mistral AI). In July 2025, just months after creation, it emerged from stealth with a $105M seed round co-led by Eclipse and Khosla Ventures, with Bpifrance, HSG, Eric Schmidt, Xavier Niel, Daniela Rus and Vladlen Koltun also investing, to build a data engine for training AI models for robots.
The team's thesis was that robotic manipulation is the most valuable unsolved problem in robotics, and that its real bottleneck is data: human hand motions do not transfer to robot hardware, and teleoperation setups cannot scale. So Genesis went full stack rather than model-only — designing Genesis Hand, a robotic hand matching human size and soft-contact physics, developed with Chinese partner Wuji Tech, plus a lightweight sensor-loaded glove and egocentric video capture so people can collect demonstrations while doing their normal jobs, and a simulation system to speed up evaluation.
In May 2026 Genesis released GENE-26.5, its first foundation model, trained across language, vision, tactile sensing, proprioception and action. A single set of model weights executed long-horizon real-world tasks at 1x speed — cooking with egg cracking and knife work, lab pipetting, smoothie preparation, wire harnessing and a bimanual Rubik's cube solve — using under an hour of task-specific robot data per skill. Eric Schmidt called it a milestone for the team and the industry; the company was hiring across Paris, California and London and planning a first full-body general-purpose robot.
What has to be true
- Over 80% of physical labor is manipulation and almost none of it has been recorded digitally, so data scarcity — not model architecture — is the bottleneck.
- Closing the embodiment gap in hardware, with a 1:1 human-sized hand and soft-contact physics, makes human demonstrations transfer to robots without retargeting.
- A wearable glove plus egocentric video means real jobs become data collection, unlocking a scale teleoperation setups cannot reach.
- Going full stack removes dependence on third-party hardware that constrains what data can be collected, letting model and hand co-evolve.
What can be applied
If data is the bottleneck, don't wait for someone else's hardware: control the full stack — model, hand, glove and simulator — so real work becomes training data.
Aftermath
As of 2026-05-07, Genesis AI had publicly released GENE-26.5 and was planning to reveal its first general-purpose full-body robot shortly. The startup had ~60 employees across offices in Paris, California and London and was in talks with customers, positioning the data glove for fields like pharma and manufacturing; co-founder Gervet acknowledged open questions about whether workers would wear the sensors and how data sharing would be handled.
Sources
- Khosla-backed robotics startup Genesis AI has gone full stack, demo shows
- GENE-26.5: Advancing Robotic Manipulation to Human Level
- Ex-Mistral employee raises $105m for AI and robotics startup
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