Generalist AI GEN-1.5

2026-08-19 · Publisher : Generalist AI · Domain : Robotics

Generalist AI unveils GEN-1.5, a model that learns a new physical task from a single 3-to-12-second demonstration without weight updates, or with 1 to 10 gradient steps on minutes of data.

What changes

One-shot learning through a 3-to-12-second "physical prompt".
No weight updates needed.
Adaptation possible with 1 to 10 gradient steps on minutes of data.
Action trajectories output at 100 Hz.
Generalist speaks of a "GPT-3" moment for robotics, on still-simple tasks.

### Background On August 19, 2026, Generalist AI unveiled GEN-1.5. According to the company, it is the first model in which one-shot and few-shot learning of physical skills emerges at scale: it learns a new task in seconds from a single demonstration, without fine-tuning. ### Specifications - **"Physical prompting"**: the model learns a new task when given a single 3-to-12-second demonstration in context, with no training. - **Compositional generalization**: given two different demonstrations in context, the model can chain tasks. - **Fast adaptation**: otherwise, 1 to 10 gradient steps on minutes of data suffice. - **Outputs**: 100 Hz action trajectories from vision, language and proprioception. - **Other capabilities**: zero-shot sim-to-real transfer and in-context learning from human demonstrations. - **Limits**: the company notes the tasks remain simple and short-horizon. ### Why it matters Generalist draws a parallel with GPT-3, which made few-shot learning a core capability of language models. Applied to robots, this leap would let people show a task rather than program it or collect hours of data. ### Key takeaway GEN-1.5 landed the day after Skild AI's S1: in-context learning became the new frontier of robot models in summer 2026.

Generalist AI GEN-1.5

What changes

Links