Generalist AI has introduced GEN-1.5, a robot foundation model the company says can pick up a new manipulation task from three to 12 seconds of a single demonstration placed directly in its context window, with no additional training. Generalist calls the technique "physical prompting" the sensorimotor equivalent of feeding a language model a one-shot example, and says the ability emerged from large-scale pretraining rather than being explicitly built in.
Across 10 diverse tasks, including handling zippers, opening jars, and retrieving money from a wallet, GEN-1.5 averaged 59 percent success from a single in-context demonstration with zero gradient updates, a result the company reports with a ±10 percent standard deviation. Moving from prompting to light fine-tuning, 10 gradient steps on five minutes of data per task, roughly 50 demonstrations, raised the average to 83 percent. Generalist notes that in some cases the no-training in-context result matched or beat a few gradient steps on the same data.
The number a manufacturer should sit with is not the success rate but the data budget. Adding a task to a robotic cell has traditionally meant writing new code or collecting a full training run, often tens of thousands of gradient steps. GEN-1.5 does it from seconds of demonstration or, at most, a few minutes of data and a handful of steps. Engineers running high-mix lines, where new parts, fixtures, and tool changes arrive constantly and reprogramming is the bottleneck, are the ones for whom that shift in setup cost matters most.
Generalist also reports that the model improvises rather than fails when conditions leave the demonstration. After being fine-tuned on five minutes of human demonstrations of sweeping a block into a bowl with a brush, the model was handed tools it was not trained to use. Given a dustpan, it did not attempt the brushing motion but composed a new sequence, lifting the block with the dustpan and tipping it into the bowl. Handed a banana, it used the banana as a makeshift brush. The company says neither the fine-tuning data nor, as far as it knows, the pretraining data contained a dustpan used this way.
The caveats are Generalist's own. The tasks are simple and short-horizon, the success rates are modest, and the tool-improvisation examples are demonstrated cases rather than benchmark averages. The blog names no deployment partners, hardware compatibility, or pricing, and says nothing about inspection or quality work, so how physical prompting holds up across a broad range of production tasks remains untested outside the lab. GEN-1.5 has trained continuously for more than eight months, and Generalist says it has not yet found where the curve flattens.
