Generalist AI has released GEN-1.5, a robot foundation model that can learn and carry out new physical tasks from a single 3-12 second demonstration, according to a Techub summary citing MarkTechPost. The company said the system does not require gradient updates, fine-tuning, or task-specific programming to pick up a new task.
Single-demo performance across 10 manipulation tasks
In 10 different manipulation tasks, the model, without any prior training for those tasks, posted an average success rate of 59% from a single in-context prompt. With five minutes of data for each task, or about 50 demonstrations, followed by 10 gradient update steps, the success rate increased to 83%.
"Physical prompting" within a 30-second context window
GEN-1.5 uses what the report described as a "physical prompting" mechanism. Through a drag-and-drop interface, users can place sensorimotor examples into the model's 30-second context window, and the model then performs the task.
The report said the model's capabilities came from eight months of continual pretraining on physical interaction data collected across homes, warehouses, and factories, rather than from hand-crafted task design.
Research version only, with no public weights or API
GEN-1.5 is currently a research version. There are no public model weights, no API, and no self-serve product at this point. Access is available only through direct partnerships.
Simulation-to-real transfer and human imitation
The model also showed combinatorial generalization, zero-shot sim-to-real transfer, and human imitation, according to the report. One example said demonstrations recorded in simulation could be used directly to prompt a real robot. In another case, after limited fine-tuning, the model was able to use a banana as a temporary brush to complete a brushing task.