Generalist AI is a company focused on creating general intelligence for the physical world. Its stated aim is to develop systems that allow robots to become useful after brief human input rather than lengthy specialized programming.
GEN-1.5 is the latest robot foundation model from the company. This one takes video from the robot’s cameras along with other sensor readings, spoken or written language, and proprioceptive signals, which are the robot’s own sense of its joint positions and movements. It then produces continuous action commands at one hundred times per second.
The central result is that the model can acquire a new simple skill for brief tasks after seeing only one demonstration lasting three to twelve seconds. The demonstration is placed into a thirty-second memory window, an approach called physical prompting or in-context learning. In-context learning means the model uses the new example stored in its temporary memory without changing its internal weights. Across ten basic tasks such as twisting a jar lid, unzipping a pencil pouch, or retrieving money from a purse, the average success rate reached fifty-nine percent. These tasks remain simple and brief.
How the model adapts further
When a small amount of additional practice data is supplied, performance rises. After ten small adjustment steps, performed on roughly five minutes of demonstrations, average success climbed to eighty-three percent. The same model can also chain two separate demonstrations into one longer sequence, transfer a simulated demonstration to a real robot even though no simulation data appeared during pretraining, and in limited cases copy a movement shown by a human hand. It sometimes invents alternative strategies or uses unfamiliar tools such as a dustpan when the original tool is unavailable.
All of these behaviors appeared after more than eight months of continuous pretraining on large volumes of real physical interaction data collected in homes, warehouses, and factories. No special architecture changes, meta-learning loops, or extra training goals were added to encourage one-shot learning. Success rates are still modest and the skills remain more fragile than those obtained by conventional fine-tuning. But this can be seen as an early milestone showing that, past a certain scale of physical pretraining, adaptation to new tasks can become far less costly.