Generalist’s GEN-1 foundation model now supports a range of robot end effectors

Generalist has expanded its GEN-1 embodied foundation model to support a broad range of robot end effectors, including five-fingered hands, specialized tools, off-the-shelf devices, printed components, and modifications of its standard two-finger grippers. The company said its dataset now includes more than 500,000 hours of real interaction data and about 9,000 gripper variations. These end effectors were selected to expose the model to diverse contact physics, actuation methods, camera positions, and tool geometries.

According to Generalist, training across many physical interfaces is intended to help GEN-1 learn sensorimotor representations that transfer between different ways of grasping, pushing, pulling, twisting, and contacting surfaces. The article cites examples such as screwdrivers, tape dispensers, tongs, spatulas, scrapers, box cutters, peelers, and whisks. These tools impose differing requirements, including tension control, compliance, distributed surface contact, constrained force paths, and perception of thin structures.

The company analyzes how fine-tuning GEN-1 for a new tool changes model weights, treating the difference from the pretrained model as a task update. It examines changes in sensor processing, harmonic reasoning, and actuation components to identify where a tool introduces novelty. Generalist said whisk fine-tuning alters sensor-processing weights more than peeler fine-tuning, suggesting a need for additional data involving thin or visually sparse tools.

Generalist also tested swapping an end effector while the model was running a task. It said the unchanged model perceived the replacement tool and selected a different trajectory and contact strategy to pursue the same goal. The company argues that interchangeable robot tooling, combined with a model able to recognize and adapt to it, could allow robots to select specialized tools when appropriate.