On May 6, Genesis AI, a robotics company split between Paris and San Carlos, California, introduced its first model, GENE-26.5.
The name is straightforward: Genesis robotic Neural Engine, versioned for May 2026. What mattered more was the demo: a robotic hand performing work that usually still belongs to human fingers.
What the hand can do
The company did not show a simple factory pick-and-place routine. Its video centered on fine manipulation:
- a 20-step cooking sequence, including cutting a tomato, cracking an egg and frying it;
- two-handed coordination to make a milkshake;
- pipetting in a laboratory setting;
- connecting cables;
- solving a Rubik's Cube;
- grasping four objects at once;
- playing piano, which the video described as close to human level.
Those are the tasks robotics companies have been careful not to claim too loudly over the past three years. Boston Dynamics' Atlas, Figure's 02 and 1X's Neo can walk, stand and move boxes. None of them is known for frying an egg.
Fine manipulation remains one of robotics' hardest unsolved problems.
Who is behind Genesis AI
CEO Zhou Xian came out of CMU. President Théophile Gervet is a former Mistral AI researcher. The 60-person team is spread across Paris, California and London, with roughly 40% to 45% in Europe and 50% to 55% in the United States.
The company has raised a $105 million seed round, co-led by Eclipse and Khosla Ventures. Other backers include Bpifrance, HSG and a group of individual investors, including former Google CEO Eric Schmidt, Xavier Niel, former MIT CSAIL director Daniela Rus and Apple AI research scientist Vladlen Koltun.
This is not a casual investor list. It is packed with people who understand the robotics and AI track well.
Schmidt called it "a paradigm shift in robotics."
The key is not only the model. It is the data glove.
The bottleneck in robotics has rarely been only algorithms. It has been data.
A human baby spends months learning how to grasp objects, while the brain processes visual, tactile and proprioceptive signals at enormous scale. Robots do not have that biological sensor stack. They usually rely on human teleoperation, where a person wearing a VR headset slowly demonstrates motions to a robotic arm.
That approach has two fatal limits: it is slow, with single actions taking hours to collect, and it is expensive, with teleoperation systems often starting in the hundreds of thousands of dollars.
Genesis AI built a data-collection glove instead. The glove is packed with tactile-sensing electronic skin and is designed to mimic the form and function of the human hand. Workers can wear it in real factories and real kitchens, turning their motions directly into training data.
The numbers are the important part:
- the glove hardware costs 1/100 of the industry-standard setup;
- data collection is five times more efficient than conventional teleoperation.
In plain terms, Genesis wants a company to spend something closer to the price of a Model 3 and build a thousand-person data collection workforce, gathering demonstrations across many settings every day at a scale old teleoperation methods could not reach.
That is the core of the company's valuation story.
How Genesis differs from Physical Intelligence and Skild AI
Three foundation-model companies now define the front of the robotics race:
| Company | Valuation / funding | Route |
|---|---|---|
| Physical Intelligence | $4.4 billion valuation in 2025 | Software only, no hardware |
| Skild AI | $1.5 billion valuation in 2024 | General robot brain, hardware-agnostic |
| Genesis AI | $105 million seed round | Full stack: model, hardware, simulation and data |
Physical Intelligence and Skild AI are betting that the strongest robot brain can plug into many bodies. That is a story about generality.
Genesis is making the opposite bet: it wants to build the brain, the hand, the data-collection device and the simulation environment itself.
That may look slower in the short term because hardware burns capital. Over the long run, Genesis may get further if the tight coupling between software and hardware keeps forcing software-only companies to retune their systems for each new machine.
Khosla's Vinod Khosla said foundation models and simulation capabilities "are poised to dramatically increase development speeds."
Why it matters
The Physical AI story has had an awkward gap over the past two years: software models advanced quickly in 2024 and 2025, but robots were still mostly at the stage of standing and walking.
Figure, 1X and Boston Dynamics keep producing sharper demos. In real production environments, however, the work that lands is still largely material handling, inspection and reception, jobs traditional robotics could already address.
Fine manipulation will probably require either the next breakthrough in VLA models, or brute-force growth in data scale.
Genesis AI chose the second path: use cheaper hardware, human demonstration and simulation to build one of the largest high-quality data engines in the sector.
The bet is simple: when model architectures start to look similar, whoever owns more high-quality data wins.
That logic has already been tested once in LLMs. Robotics is where it gets tested next.
Sources: Khosla-backed robotics startup Genesis AI has gone full stack, demo shows (TechCrunch); CocoLoop; Genesis AI Unveils GENE-26.5, the First AI Brain to Enable Robots with Human-Level Physical Manipulation Capabilities (PR Newswire)