A robotic hand cracks an egg. Just one hand—the other remains still. The fingers curl with surprising delicacy, tapping the shell against a countertop, pulling it apart, depositing the contents cleanly. Then the hand reaches for a dish towel and wipes away the residue.
It's the kind of mundane task humans perform without thinking. For robots, it's been maddeningly difficult.
The demonstration comes from Genesis AI, a Paris-based startup that surfaced from stealth in July 2025 with $105 million in seed funding and ambitions that sound, frankly, audacious. The company says it's building the foundation for general-purpose robotics—not just the software or just the hardware, but the entire stack. The egg-cracking video, released in early May alongside clips of a robot hand playing piano and assembling wire harnesses, suggests they've made progress. Whether that progress means anything outside a laboratory is another matter entirely.
The Full-Stack Gamble
Genesis unveiled its GENE-26.5 foundation model and proprietary "Genesis Hand 1.0" on May 6, a strategic pivot for a company that originally pitched itself as an AI software provider for existing robotic platforms. Now it's doing everything: the model, the hand, even the high-fidelity physics simulation environment that trains them.
This is a riskier bet. Building hardware is expensive, slow, and littered with the wreckage of well-funded startups that underestimated the challenge. But Genesis co-founder Zhou Xian—who earned his robotics PhD from Carnegie Mellon—argues the company needs to control the full experience. "You can't optimize what you don't control," he told TechCrunch, explaining why Genesis designed its own hand in partnership with Chinese manufacturer Wuji Tech.
The demos are impressive, if you're inclined to be impressed. A 20-step cooking sequence. A Rubik's Cube solve. Lab pipetting that requires precision down to microliters. These aren't simple pick-and-place operations. They demand vision, planning, tactile feedback, and the kind of fine motor control that has stymied roboticists for decades.
Genesis claims its sensor-rich data collection glove enables "1:1 skill transfer"—a human performs a task while wearing the glove, and the robot learns to replicate it. The company frames this as the solution to robotics' fundamental data problem: unlike language models that can feast on the internet, robots need physical demonstrations, and those are scarce and expensive to collect.
Perhaps. Or perhaps Genesis has simply found a clever way to generate training data in controlled settings that won't generalize when a robot encounters a kitchen that doesn't look like the test lab.
The Money and the Minds
Eclipse Ventures and Khosla Ventures co-led Genesis's seed round in July 2025—an unusually large seed, even by the inflated standards of AI fundraising. Other backers included Bpifrance, HongShan (the renamed Sequoia China), and marquee angels Eric Schmidt and Xavier Niel. Bloomberg noted at the time how rare it was to see American and Chinese venture capital collaborating on the same cap table, a testament to either Genesis's cross-border appeal or the investors' fear of missing out on the next big robotics play.
Zhou Xian, the CEO, brings academic credibility—he led development of the Genesis physics simulation platform during his doctoral work. His co-founder, Théophile Gervet, came from Mistral AI, where he contributed to research on the Mixtral family of models. The founding team launched the company in December 2023, a compressed timeline for what they're attempting.
As of May 2026, Genesis employs about 60 people across offices in Paris, San Carlos, California, and a newly opened London location. That's up from roughly 20 researchers at the seed announcement. In March, the company hired Vivian Sun—who previously held roles at Amazon, autonomous vehicle startup Waabi, and self-driving truck company TuSimple—as VP of Commercial & Strategy. Sun's pedigree suggests Genesis is moving beyond research demonstrations toward actual deployments, though the company has been vague about customer timelines.
Not Alone in the Arena
The robotics foundation model race has attracted an almost comical amount of capital lately, most of it chasing the same bet: that AI models can finally crack the code on physical manipulation.
Physical Intelligence pulled in $400 million in November 2024 with backing from Jeff Bezos and OpenAI. Skild AI reached a $14 billion valuation this January after raising roughly $1.4 billion. Figure AI—known for its humanoid robots—secured $675 million in early 2024 at a $2.6 billion post-money valuation, with a roster of investors including Bezos, Nvidia, Microsoft, and OpenAI. The pattern repeats: eye-watering sums flowing toward startups promising to solve one of AI's hardest remaining problems.
Genesis is smaller, younger, and less capitalized than several of its competitors. But it's made a different architectural choice. Where others are building software that works across multiple robot platforms, Genesis is controlling the hardware, betting that vertical integration will let it iterate faster. The hand demonstrations are a proof point. In its May press release, Genesis teased plans to reveal "its first general-purpose robot" soon—what Xian described to TechCrunch as "a full-body robot, not just hands."
The company is already in discussions with potential customers and plans to deploy its data collection gloves in workplace settings to gather training data at industrial scale. Whether those workplaces materialize remains to be seen.
The Reality Check
Here's what Genesis has actually proven: its robots can perform impressive tasks in structured environments with good lighting, clean surfaces, and objects positioned just so. The videos look polished. They should—Genesis has world-class researchers and more than $100 million to burn.
The question, as always with robotics, is generalization. Can a robot trained on lab pipetting handle a different brand of pipette in a hospital pharmacy? Can the hand that cracks eggs in a test kitchen do the same thing at 6 a.m. in a restaurant with flour on the counter and a line cook yelling in the background?
Training AI to manipulate physical objects requires vastly more data than language models, and that data is expensive and time-consuming to collect. Genesis's simulation environment and teleoperation glove are meant to shortcut that process, generating synthetic training examples and capturing human demonstrations efficiently. If that approach works at scale, the company has a real advantage. If the sim-to-real gap proves wider than Genesis expects, the demos become less meaningful.
The startup was founded just over a year ago. Its competitors have deeper war chests and, in some cases, years of head start. But Genesis has moved quickly—from founding to major funding to hardware release in roughly 17 months. The May demonstrations suggest the team has built something functional. Now comes the harder part: proving it works outside the lab, in places where the lighting isn't perfect, the task changes without warning, and nobody's standing by to reset the scene when something goes wrong.
That's where robotics companies tend to stumble. Genesis may turn out differently. Or it may discover, like so many before it, that the gap between demo and deployment is wider than $105 million can bridge.
