TechCapsules

AI-summarised brief · reviewed before publication

Generalist has updated its GEN-1 embodied foundation model to support a diverse array of robot end effectors, ranging from five-fingered hands to specialized tools. The model leverages over half a million hours of real interaction data, including approximately 9,000 variations of grippers and custom modifications. This expansion allows GEN-1 to learn universal sensorimotor representations that transfer across radically different physical interactions. By training on varied form factors with distinct actuation schemes, the model develops general physical commonsense regarding geometry, friction, and dynamics. Generalist compares this approach to multilingual language models, where learning across different "physical languages" improves overall reasoning capabilities. The company analyzes weight shifts during fine-tuning to quantify how novel each end effector is, identifying which tools provide the most significant learning signals. This method helps separate tool-specific behaviors from universal physical laws, enabling the robot to switch between tools effectively. The initiative aims to enhance physical intelligence by exposing the model to a broad spectrum of contact physics and interaction dynamics through deliberate dataset expansion and rigorous benchmark evaluation.