GENE-26.5 Gives Hardware-Agnostic Robots Human-Scale Dexterity

On May 6, 2026, French startup Genesis AI shifted to a full-stack robotics strategy with the launch of its first foundation model, GENE-26.5. Founded by researchers from Carnegie Mellon and Mistral AI, the company paired the software release with a custom-designed human-scale robotic hand. The system is designed to execute dexterous manipulation and long-horizon tasks across different hardware platforms.

The GENE-26.5 Foundation Model

GENE-26.5 operates as a hardware-agnostic robotics foundation model. Unlike tightly coupled proprietary systems, Genesis AI engineered the model to run on robots manufactured by third parties. This allows industrial operators to upgrade the software intelligence of existing fleets without overhauling their hardware footprint.

The model targets tasks that require high-precision manipulation rather than simple locomotion. Early benchmarks and demonstrations indicate that the system can solve spatial problems using in-air manipulation, a capability historically difficult to program via hardcoded rules.

Closing the Embodiment Gap

To train GENE-26.5, Genesis AI built a proprietary data engine centered around a sensor-loaded glove. Human workers wear the glove to perform daily physical tasks, generating high-quality real-world data at scale. This telemetry feeds into a closed-loop system integrated with a high-fidelity physics simulation engine capable of generating synthetic training data at 430,000 times faster than real-time.

The company also designed a custom robotic hand featuring a 1:1:1 mapping with human anatomy. Traditional two-finger grippers often struggle with complex tasks, creating an embodiment gap where the model understands the task but the hardware cannot physically execute it. The anatomical mapping ensures that the skills learned from human demonstrations translate directly to the physical robot, expanding the range of feasible untrained tasks in industrial environments.

Demo Capabilities

Genesis AI released autonomous, non-teleoperated demonstrations of GENE-26.5 running at 1x speed. The system executed a 20-step bimanual cooking sequence, which included chopping tomatoes and cracking an egg with one hand.

For technical dexterity, the model solved a Rubik’s Cube, played the piano at 130 beats per minute, and performed high-precision lab tasks such as pipetting. The system also completed wire harnessing, arranging and securing complex wire bundles. This specific task is highly sought after in electronics manufacturing due to its historical resistance to automation.

Strategic Focus and Hardware Roadmap

Backed by a $105 million seed round led by Khosla Ventures and Eclipse Ventures, Genesis AI is explicitly targeting the European market. The company aims to support industrial reindustrialization across France, Germany, and Italy, focusing on the automotive, electronics, and pharmaceutical sectors.

Genesis AI currently operates with approximately 60 employees split between San Carlos and Paris. The company plans to reveal its first complete general-purpose robot, reported to be a wheeled humanoid with two arms, in the near future.

If you manage industrial automation, hardware-agnostic models like GENE-26.5 alter your procurement cycle. You can now separate the evaluation of AI inference software from the physical chassis, allowing you to deploy updated manipulation policies onto existing robotic platforms.