Generalist AI has unveiled GEN-1.5, a robot foundation model that learns new physical tasks from a single demonstration. By feeding 3–12 seconds of sensorimotor data into its 30-second context window, the robot can execute the task immediately—no gradient updates, no fine-tuning, and no task-specific programming required.
In testing across 10 diverse manipulation tasks, GEN-1.5 achieved an average success rate of 59% (±10% standard deviation) from the pretrained model alone, relying purely on one-shot in-context prompting. When given ten gradient steps on five minutes of task-specific data, performance improved to 83% (±9%). The company refers to this mechanism as "physical prompting," emphasizing that it was never explicitly trained for this capability. There were no architectural changes, meta-learning loops, or auxiliary objectives—the skill emerged from over eight months of continuous pretraining on physical interaction data.
While the tasks evaluated are simple and short-horizon—a point Generalist AI openly acknowledges—this marks a notable milestone: it is the first model the team knows of where one-shot learning of physical skills has emerged at scale.
Is it deployable?
Not yet. This is a research release. There are no public weights, no API, no pricing page, and no self-serve product. Generalist AI runs GEN-1.5 on its own fleet and data engine, and access is currently granted only through direct partnerships. As of 2026, this positions GEN-1.5 as a glimpse into the future of adaptable robotics—where foundational models could soon enable rapid, flexible automation without custom training—rather than a ready-to-use commercial solution.
via MarkTechPost