Generalist AI has released GEN-1.5, a research-oriented robot foundation model capable of learning new physical manipulation tasks from a single 3–12 second demonstration via its 30-second context window. This capability, termed "physical prompting," emerged spontaneously from over eight months of continuous pretraining on physical interaction data rather than being explicitly designed.
- One-shot in-context prompting achieved an average 59% success rate across 10 diverse tasks without any gradient updates or fine-tuning.
- Applying ten gradient steps on five minutes of data per task raised the success rate to 83%, with weight changes under 0.15%.
- The model demonstrated compositional generalization, zero-shot sim-to-real transfer, and human-to-robot imitation capabilities that were not explicitly trained for.
The release serves as a research signal regarding the scaling of physical prompting, though it is not yet deployable as there are no public weights, API, or self-serve product available.