Zero-shot generalization

Zero-shot generalization is the ability of a learned policy to perform tasks it was never trained on, by transferring from related training distributions. The strict version: zero gradient updates on the target task — the policy evaluated at test time is identical to the policy at end of training. Looser versions: few-shot fine-tuning on a handful of demonstrations, prompt-conditioning that effectively reweights training-distribution behaviors, or test-time adaptation. Each is a meaningfully different capability claim.

The distinction matters because zero-shot claims routinely get conflated with few-shot ones in physical-AI marketing. Physical Intelligence's π0 paper makes the distinction carefully; some other programs are less precise. When a model is described as "zero-shot" without surrounding methodology, the safe reading is that some form of few-shot adaptation was used. See also: sim-to-real transfer, which has the analogous strict-vs-loose split for the simulation-to-hardware gap.