Genesis AI · 2026-05-06 · major
Genesis AI GENE-26.5 — Foundation Model + Human-Scale Hand for Cooking, Piano, and Rubik's Cubes
Khosla- and Eclipse-backed Genesis AI debuts GENE-26.5, a robotic-manipulation foundation model paired with a human-scale dexterous hand demoed cracking eggs, slicing tomatoes, playing piano, and solving a Rubik's cube.
An AI 'brain' for robot hands trained on human-scale data, demoed cooking a 20-step meal and solving a Rubik's cube.
Key specs
| Seed funding | $105M |
|---|---|
| Hand cost vs. typical | 100x cheaper |
| Data efficiency vs. teleop | 5x |
What is it?
GENE-26.5 is Genesis AI's first foundation model for robotic manipulation, paired with a custom-built robot hand whose form factor matches a human hand. The Paris- and Palo Alto-based startup, founded by Zhou Xian and former Mistral AI researcher Théophile Gervet, raised $105M in seed funding in 2025 from Eclipse and Khosla.
How does it work?
Training data comes from four overlapping sources: a sensor-loaded data-collection glove with 1:1 human-hand mapping, egocentric video from human workers, internet video, and a proprietary high-realism simulator that narrows the sim-to-real gap. The matching robot hand is reportedly 100x cheaper to manufacture than typical research hands and yields 5x the data per teleop session, because the same gloves humans wear can drive both training and capture.
Why does it matter?
Most robot foundation models train on parallel-jaw gripper data because dexterous-hand data is scarce. Genesis is betting that owning the full stack — hand hardware, data-collection glove, simulator, and model — is the only way to scale manipulation training data. The launch video shows multi-step cooking, piano playing, lab pipetting, and a Rubik's cube solve, putting Genesis on the same demo board as Physical Intelligence, Figure, and 1X.
Who is it for?
Embodied AI researchers, robotics labs, humanoid startups
Try it
Demo video: https://youtu.be/6K_bGH54ltI