The robotics specialist Generalist AI presented the new Gen-1 model, which represents an important milestone on the way to a universally applicable physical AI for practical tasks. The system is designed as an ‘Emboded Foundation Model’ and was trained on the basis of large-scale data from real interactions. According to the company, Gen-1 achieves a success rate of 99 percent for certain tasks and executes them up to three times faster than previous systems. In addition, the model is considered to be highly data-efficient: it only takes about an hour of robot-specific data to adapt to new tasks. Demonstrations show robots that autonomously perform tasks such as packing items or assembling components over long periods of time. The model is based on the predecessor Gen-0 and expands its approach with more computing power as well as new training and inference methods. A central aspect is the use of extensive data of human activities (“pre-training”) instead of exclusively relying on expensive teleoperation data sets. Despite the advances, the company admits that not all tasks provide production-ready performance and further improvements in speed and reliability are needed for a wide application.
via roboticsandautomationnews.com