Generalist's GEN-1.5 learns new robot tasks from seconds of demos, zero fine-tuning

量子位 · wechat · 2026-08-21

Generalist AI released GEN-1.5, a robot foundation model with one-shot in-context learning: show it a 3-12 second demonstration (a "physical prompt") and it performs the new task with zero gradient updates or fine-tuning, in roughly half a minute. It can also chain two separate demos into one continuous task and transfer a simulator demo straight to the real world.

Notably, the ability wasn't explicitly engineered—no special architecture, no meta-learning loop—it emerged from large-scale pretraining on continuous real-world physical data, echoing how in-context learning emerged in GPT-3. During pretraining the team saw required fine-tuning data shrink from hundreds of steps to tens, then 10, then zero.

Benchmarks: across 10 tasks, 59% average success with zero gradient updates; 83% with 5 minutes of extra data and 10 gradient steps per task. Current tasks are mostly short-horizon and stability still lags full fine-tuning, but 59% with no training is what excites researchers.

Original post →

Generalist's GEN-1.5 learns new robot tasks from seconds of demos, zero fine-tuning

More from Embodied