Aran Komatsuzaki· @arankomatsuzaki · X·· 2 小时前AI 评分60
AI 导读
LiberAI 发布世界动作模型 Liber-0，基于 220 万小时以上人类经验预训练，在 60 个任务、600 次试验中零样本整体成功率为 36.0%。作者指出，动作 grounding 与具身对齐让模型把人类经验转化为机器人可执行的动作，其中抓取放置任务成功率 63.5%，其他操作任务 34.0%，多步任务 10.5%。作者认为人类经验数据比机器人数据更易获取，若该路线成立，可让训练数据采集更快更便宜，新任务不必每一步都依赖机器人参与。
正文 · 原文
LiberAI just released Liber-0, a WAM pretrained on 2.2M+ hours of human experience.
With zero robot demos for the target task, it reached 36.0% success across 60 tasks and 600 trials. Action grounding and embodiment alignment let it turn human experience into actions a robot can execute.
Human experience is much easier to get than robot data. If this holds up, it could make collecting training data a lot faster and cheaper, and you wouldn't need a robot in the loop for every new task.
Introducing Liber-0, our World Action Model for zero-shot robot manipulation—including dexterous, multi-step tasks. Across 60 tasks proposed by GPT-6 and 600 trials, Liber-0 achieved 36% overall success without robot demonstrations collected specifically for those tasks: • Pick-and-place: 63.5% • Other manipulation tasks: 34.0% • Challenging multi-step tasks: 10.5% Zero-shot capability extends beyond individual actions to coordinated sequences. Pretrained on 2.2M+ hours of human experience, Liber-0 jointly predicts future observations and actions to turn human experience into robot skills. Technical blog: https://www.liberai.com/blog/liber0在 X 查看被引用的帖子
来源：Aran Komatsuzaki · x.com