Company Generalist AI has introduced the GEN-1.5 robotics model, which can learn physical actions in just 3 to 12 seconds after a single demonstration. Engineers do not need to retrain it for each new task: the model watches an example and immediately tries to repeat the action.
In tests, GEN-1.5 learned to open a book, remove a lid from a jar, stack cups, sweep away debris, and take money out of a wallet. Across ten tasks, the average success rate after one demonstration was 59%. With five minutes of additional demonstrations and a small amount of fine-tuning, the figure rose to 83%.
Its creators call the demonstration a “physical prompt” — in effect, a set of instructions through action. This approach could make robots much more flexible: instead of lengthy reconfiguration, they would just need to be shown what to do.
