Robotics company Generalist has introduced GEN-1.5, a foundation model designed to learn a new physical skill after seeing only one short demonstration. The company describes the approach as an early step towards robots that can be shown what to do on the spot instead of being retrained for every new object or task.
In Generalist’s demonstrations, an example of the desired movement is placed directly into the model’s working context. GEN-1.5 then attempts the task immediately, using feedback from the robot’s sensors to adjust its actions as conditions change. The company calls this “physical prompting,” drawing a comparison with the way language models learn a pattern from examples included in a prompt.
One demonstration, no training step
The key difference is speed. Traditional robot learning often requires collecting many demonstrations and updating a model before it can perform a new task. Generalist says GEN-1.5 can learn from an example lasting only a few seconds, without gradient updates or task-specific fine-tuning.
Across a set of 10 varied tasks, the company reports an average one-shot success rate of 59 per cent. When given roughly five minutes of demonstration data and 10 training steps, the reported success rate rose to 83 per cent. The model emerged after more than eight months of continuous pretraining on physical-interaction data.
Combining skills and crossing from simulation to reality
GEN-1.5 can also combine separate examples. Generalist says the model can watch demonstrations of two different behaviours and connect them into a longer sequence, including movements that were not explicitly shown. This type of compositional learning could eventually make it easier to build useful workflows from smaller skills.
The company also reports zero-shot transfer from simulation to a physical robot. In that test, a simulated demonstration was used as the prompt even though simulation data was not included in the model’s pretraining. GEN-1.5 can additionally learn from human demonstrations, pointing towards more natural ways for people to teach machines.
Promising results, with important limits
Generalist is careful to describe the demonstrated tasks as simple and short-horizon. A 59 per cent average success rate also means one-shot performance remains unreliable for many real-world uses, especially where mistakes could damage equipment or create safety risks. The published results come from the company and will require broader testing across robots, environments and longer tasks.
Even with those limits, GEN-1.5 points to a significant change in how robots may be programmed. If machines can reliably learn from a handful of demonstrations, businesses could adapt them to new products and processes more quickly, while non-specialists could teach tasks without writing code or building a new training pipeline.
Sources: Generalist launch video; Generalist GEN-1.5 research article.