GENE-26.5 is an AI Brain That Enables Robots to Crack Eggs and Play Piano with Human Hands

Genesis AI just released a new AI model called GENE-26.5 that lets robots handle everyday objects with the same ease and precision people use. In tech demos released this week, one robot cracks an egg with one hand without spilling a drop, then uses both hands to chop tomatoes and stir a pan during a full meal. Another sorts four different items into bins using just its fingers, while a third solves a Rubik’s Cube in mid-air with quick wrist turns and steady grips.

The model comes from a young robotics company based in Paris and California. Genesis AI raised 105 million dollars last year and focuses on building complete systems that combine software and hardware. Its founders explain that most robots still struggle with tasks that require careful touch and coordination because they lack the right data and physical setup. GENE-26.5 changes that by learning directly from human movements recorded in real environments.

At the heart of the system is a robotic hand designed to match the size and shape of a human hand. The Genesis Hand 1.0 has twenty moveable joints that bend easily and employs soft material on the palm and fingers to mimic how skin and muscle interact with items. People wear a unique glove during typical work that captures every finger move and contact pressure. The glove is one-hundredth the cost of older equipment and collects information five times faster while providing cleaner signals. Because the glove, human hand, and robot hand all line up perfectly, skills flow directly from person to machine without the normal translation issues.

GENE-26.5 is based on more than 200,000 hours of recorded human activities. Some footage shows people using the glove while cooking, assembling parts, or doing laboratory testing. Other clips show first-person perspectives from cameras installed on workers’ heads, and still more are taken from everyday footage on the internet. The model integrates linguistic instructions, camera images, joint locations, and touch signals into a single shared understanding. It predicts full sequences of movements rather than individual stages, allowing it to manage long chains of tasks such as making a smoothie from beginning to end.

Training takes place in a detailed virtual world that simulates real physics and lighting. Before performing any physical tests, robots simulate hundreds of permutations of a task. This approach accelerates development and lowers errors when the system is deployed to actual hardware. Once trained, the same model can do new jobs with very little further experience, generally only twenty or thirty minutes of real-world scenarios.