Generalist’s new physical robotics AI brings “production-level” success rates

Generalist, a pioneering company in robotic machine learning, has unveiled GEN-1, a groundbreaking physical AI system that boasts achievement levels akin to production standards across a diverse array of physical skills. These tasks, traditionally reliant on the finesse and muscle memory of human hands, can now be performed with remarkable efficiency by GEN-1. This innovative model is designed to adapt and respond to unexpected challenges by creating new movements and integrating concepts from various sources to tackle novel problems. Building on the foundation laid by its predecessor, the GEN-0 model, which was introduced as a proof of concept showcasing the potential of scaling laws in robotic training, GEN-1 represents a significant leap forward. While large language models have harnessed vast amounts of text data from the internet to enhance their training, robotics has faced challenges due to the lack of accessible quality data regarding human object manipulation. To bridge this gap, Generalist has developed “data hands”—wearable pincers that capture minute movements and visual data as humans engage in manual tasks. The company reports that it has amassed over 500,000 hours and vast amounts of physical interaction data to refine its physical model. The outcome is a highly autonomous system capable of performing tasks with precision, such as placing money into a wallet, folding laundry, or sorting automotive parts. GEN-1 achieves a remarkable 99% success rate on intricate mechanical tasks like folding boxes and packing devices, doing so at nearly three times the speed of its predecessor. This impressive performance is realized after just one hour of adapting its pre-training to the specific robot data relevant to its functions. Historically, complex robotic systems have depended on meticulously pre-programmed motions or were limited to single-task operations with minimal variation. What distinguishes GEN-1, according to Generalist, is its ability to improvise based on prior experiences and to react to disruptions in a natural manner, even when faced with scenarios outside its training parameters.