Generalist AI releases GEN-1.5 robot foundation model that learns new skills from 3-second demos

Robotics startup Generalist AI has released GEN-1.5, a new robot foundation model built around one-shot learning: a robot that watches a 3-12 second demonstration can immediately perform a brand-new task, with zero gradient updates and no fine-tuning.

The model goes beyond single-task imitation. It can stitch two different teaching demonstrations into one continuous task, filling in transition movements that were never shown — repositioning, adjusting its grasp, switching poses, and even recovering after an error.

Demonstrations do not even have to come from a human. Scripted policies, reinforcement learning agents, and human teleoperation inside a simulator can all be fed into the model's context as physical prompts. After watching a virtual demonstration, GEN-1.5 reproduced the task 1:1 in the real world, despite never having been trained on that task in either setting.

In tests across 10 tasks, GEN-1.5 achieved an average success rate of 59% with a single physical prompt and zero gradient updates; adding five minutes of data and 10 gradient steps per task lifted success to 83%. The team also notes that current test tasks are mostly short-horizon, relatively atomic operations, and skills learned in-context are not yet as stable as those from true fine-tuning.

What excites researchers is that Generalist says it did not design a special architecture, meta-learning loop, or auxiliary objective for one-shot learning — the capability emerged on its own during more than eight months of pre-training on large-scale physical data. During pre-training, the data needed for new tasks shrank from hundreds of fine-tuning steps to dozens, then ten, and finally to just one minute of data and a single gradient step.

That is why the release is being compared to the GPT-3 moment for robotics: GPT-3's in-context learning let models pick up new tasks without retraining, and GEN-1.5 brings the same logic into the physical world, where a demonstrated action replaces a text prompt as the robot's context.

The open questions now are whether zero-training skills stay reliable on longer-horizon, more complex tasks, and whether learning a skill by watching once becomes the standard paradigm for robot foundation models.