Generalist's GEN-1 model learns to control different robot hands

Generalist announced that its GEN-1 foundation model has been extended to support a range of robot end effectors—the hands and gripping mechanisms at the end of robot arms. The company achieved this by training GEN-1 to work with new hands, demonstrating that a single base model can learn sensorimotor policies on different robots. Sensorimotor policies are the learned mappings between what a robot perceives (via cameras and sensors) and the motor commands it sends to its actuators. By extending GEN-1 in this way, Generalist shows that foundation models can abstract over hardware variations, allowing developers to leverage the same pre-trained model across robots with different physical designs rather than building separate, robot-specific models. This flexibility could streamline robotics development and deployment, enabling faster iteration and broader applicability of a single foundation model across a heterogeneous fleet of robots.

Generalist's advancement addresses a long-standing challenge in robotics: the brittleness of models trained on a single robot design. By training GEN-1 to handle multiple end effectors, the company demonstrates that foundation models—large neural networks trained on diverse data—can abstract away hardware-specific details and learn generalizable sensorimotor skills. This follows the broader industry trend of applying foundation-model architectures (proven in language and vision) to robotics, where the same base model can theoretically adapt to different morphologies without retraining from scratch. The implication is that roboticists may be able to deploy and customize robots more flexibly, reducing the engineering overhead that currently ties software tightly to specific hardware.

No comments yet. Be the first to share your thoughts!