LiberAI 发布 Liber-0 机器人模型，未训练任务上 600 次试验成功率 36%

LiberAI’s new model, Liber-0, was tested on 60 tasks without training on robot demos collected specifically for those tasks. Those were tasks GPT-6 came up with. Across 600 trials, it achieved 36% overall success: 63.5% on pick-and-place, 34% on other manipulation and 10.5% on challenging multi-step tasks. That’s not a solved case. But a robot doing new tasks without anyone demonstrating those exact tasks on the robot is a capability worth watching. Here’s why that matters to me. Language models learned from vast amounts of text. Robots face a harder problem: physical experience doesn’t come neatly packaged as actions they can execute. Collecting robot demonstrations can be slow and expensive. But people fold, cook, clean and fix things every single day. There’s an enormous amount of experience there. The problem is that human and robot bodies don’t see or move the same way. Liber-0 uses motion recorded with LiberSense Glove, alongside human video and embodiment alignment, to help brid

来源：X · Linus Ekenstam · x.com