California-based Generalist AI has launched Gen-1, a new physical AI model that enables robots to handle tasks like folding laundry, fixing other robots and stuffing cash into wallets. The model draws on human dexterity data collected worldwide to teach robots 'physical common sense.' Co-founder Pete Florence described it as a major advance for real-world robotics.
Generalist AI unveiled Gen-1 earlier this month, powering robots through a range of practical activities. Videos from the company demonstrate the model on robotic arms performing tasks such as sorting socks by color, stacking oranges into pyramids, unzipping pencil cases and plugging in Ethernet cables. Florence, the company's co-founder and CEO, noted that Gen-1 serves as the 'brain' for various robot types, including humanoids and industrial arms. 'It's a big step forward in terms of robots designed for the real world based on intelligence born from the real world,' Florence said. Pete Florence explained that unlike traditional methods using teleoperated robots, Gen-1 was trained on data from humans wearing lightweight 'data hands' distributed globally. This captured millions of interactions, including subtle force feedback, slips and recoveries that mimic human dexterity. 'That kind of data is critical for teaching robots physical common sense, the intuitive understanding and ability to adapt in real time,' Florence added. The model shows marked improvements in success rates: servicing robot vacuums in 99% of cases, up from 50% for the prior Gen-0 version; folding boxes at 99%, up from 81%; and packaging phones at 99%, up from 62%. Gen-1 also excels in improvisation, adapting to changes like using two hands for a one-handed automotive task, a capability Florence called largely absent from prior robotics.
