AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。
TL;DR Generalist is betting that large-scale physical pretraining can give robots a reusable foundation, so teaching a new task takes a short demonstration instead of a bespoke dataset. Across GEN-0, GEN-1, and GEN-1.5, its models move from scaling evidence to roughly 99% success on selected tasks to in-context physical prompting. The open question is whether those company-reported gains hold up across complete, long-horizon jobs in real deployments. A zipper is easy to explain and surprisingly hard to operate. Hold the pouch, catch the tab, pull. Those instructions omit almost everything that matters: keeping the fabric taut, noticing a snag, changing the angle, easing off before something tears. The task is a conversation with an object, conducted through tiny movements. Generalist is building models for that conversation. Its central bet is that enough physical experience can produce a reusable foundation for robot behavior. The useful question is how much teaching a new task requires once that foundation exists. Across GEN-0, GEN-1, and the latest GEN-1.5, the company has been pushing toward an intriguing interface: show a robot an example, then let it work. Read more