LiberAI launches Liber-0 world-action model for robot control
LiberAI launched Liber-0, a world-action model pretrained on more than 2.2 million hours of human-experience data, including human videos, LiberSense Glove hand-motion data and robot data. The company says the model transfers human operation skills to robot embodiments. In LiberAI’s physical-robot tests, Liber-0 achieved a 36.0% overall success rate across 600 zero-shot trials on 60 GPT-6-generated tasks, without task-specific robot demonstrations. GPT-6 Astra recorded a 5% overall success rate on the same tasks and robot, with each task tested three times. Liber-0’s reported tasks included making bubble tea, organizing objects and operating drawers.
LiberAI says Liber-0’s use of human video and glove-recorded hand motions could reduce reliance on costly task-specific robot demonstrations and support skill transfer across robot embodiments.
11 月 4 日消息,具身智能初创企业 LiberAI 正式推出世界动作模型Liber‑0。该模型依托超 220 万小时人类经验数据完成预训练,可把人类操作经验迁移到机器人本体;真机测试结果显示,其零样本综合任务成功率显著高于 GPT‑6 Astra,同时能够大幅降低对机器人真机示范数据的依赖。
LiberAI 创始人刘松铭为清华大学计算机系 00 后博士,师从朱军教授,曾获清华特等奖学金,研究方向聚焦人工智能、具身智能与世界模型。Liber‑0 融合大规模人类视频与自研 LiberSense Glove 数据手套采集的人手动作信息,目标解决传统机器人高度依赖真机示范、跨任务、跨本体泛化能力弱的行业痛点。
在真机验证环节,LiberAI 设计了难度逐级提升的 60 项测试任务,全部由 GPT‑6 随机生成,分为抓取放置 (PnP)、短程非抓取放置、复杂多步骤任务各 20 项,每项任务重复 10 轮,合计 600 次真机零样本测试,全程没有针对任务做专门机器人示范数据采集。