Generalist AI Releases GEN-1.5 Robot Foundation Model That Learns From a Single Demonstration
aiinsider ·
Generalist AI announced GEN-1.5, a robot foundation model that can learn physical tasks from a single short demonstration or a few demonstrations and generalize to new situations without task-specific training. In evaluations across ten short manipulation tasks, the model achieved 59% average success with one-shot prompting, 83% after ten fine-tuning steps using about five minutes of data per task, and 66.5% on a held-out task after one adaptation step. Tested tasks included opening jars, unzipping a pencil pouch, retrieving money from a purse, and sweeping objects. The model incorporates video, language, sensor, and robot-position inputs and maintains roughly 30 seconds of context, allowing it to combine demonstrations, mimic human hand actions, and sometimes transfer simulated demonstrations to real robots. GEN-1.5 was pretrained for over eight months on physical-interaction data from homes, warehouses, and factories, reducing the amount of task-specific data needed. The company described this as the first model exhibiting one-shot and few-shot physical skill learning at scale, though it acknowledged the tasks were simple and short-horizon.
aiinsider