Generalist's new physical robotics AI brings “production-level” success rates
SMRTR summary
A robotic hand delicately slides a twenty-dollar bill into a leather wallet, its mechanical fingers moving with the precision once reserved for human touch. This breakthrough comes from Generalist's new GEN-1 system, a physical AI that achieves 99 percent success rates on intricate tasks like folding laundry, packing phones, and sorting auto parts.
The secret lies in "data hands," wearable sensors that captured over half a million hours of human movements, recording how people manipulate objects in microscopic detail. Unlike language models that train on internet text, robots lacked quality movement data until now.
GEN-1 operates three times faster than its predecessor and demonstrates something remarkable: the ability to improvise when things go wrong. Rather than following rigid programming, it draws from its vast training to solve unexpected problems, adapting to disruptions that fall far outside its original training scenarios.
After just one hour of learning robot-specific movements, the system can transition seamlessly from assembling delicate electronics to handling fabric, marking what Generalist calls production-level performance across diverse physical skills that previously required uniquely human dexterity and muscle memory.
SMRTR provides this summary for quick context. The original article belongs to Ars Technica.
Read the original article