Generalist AI 是一家位于麻萨诸塞州剑桥的机器人新创公司，其机械手臂能在观看一段简短的教学影片后，无需针对特定任务进行专门训练，就能掌握各种任务。最令人惊叹的是，当工具被移除时，机器人能够即兴发挥——例如用簸箕代替刷子将积木拨入碗中，或在双臂机器人无法用右夹爪抓取钞票时，自动切换到左夹爪以获得更好的操作角度。

该公司由曾任职于 Google DeepMind 和 Boston Dynamics 的 Pete Florence、Andrew Barry 和 Andy Zeng 共同创立，致力于教导机器人理解物理世界的运作方式。他们设计了类似机器人夹爪的特殊手套，配备摄影机，由人类佩戴执行各种家务来收集大量高品质训练资料，并从零开始自主建构 AI 模型，而非依赖开源语言模型。

尽管该技术展现出巨大潜力，机器人目前完成所示任务的成功率平均仅约 59%，离理想的 99% 以上仍有差距，且这些技能能否泛化到所有可能的任务或场景尚不明确。不过，乔治亚理工学院和史丹佛大学的机器人专家均对 Generalist AI 的成果给予高度评价，认为他们在通用机器人模型领域处于领先地位，最接近可实际部署的阶段。

Generalist AI, a startup based in Cambridge, Massachusetts, has developed robotic arms that can learn to perform a variety of tasks after watching a short instructional video, with no task-specific training required. The robots demonstrate remarkable improvisation abilities—when a brush was removed, one robot used a dustpan to flick a block into a bowl, and a two-armed robot spontaneously switched grippers to get a better angle when it couldn't grab banknotes.

The company was cofounded by Pete Florence, Andrew Barry, and Andy Zeng, all formerly of Google DeepMind and Boston Dynamics. Their approach centers on teaching robots intuitive physics, collecting large-scale physical interaction data through specially designed camera-equipped gloves worn by human workers. Unlike some competitors, Generalist has built its AI models entirely from scratch rather than relying on open-source language models.

While the technology shows enormous potential—particularly for manufacturing applications—the robots currently complete demonstrated tasks only about 59 percent of the time on average, far short of the ideal 99-plus percent success rate. Nonetheless, robotics experts from Georgia Tech and Stanford have praised Generalist's execution, calling them the closest to a deployable general-purpose robot model and noting that their large-scale data collection strategy appears to be paying off.