← Azimuth Daily
Capturing data…
Capturing data…
THU, 5 NOV · 70 ITEMS
"LiberAI unveils Liber-0 world action model for zero-shot robot manipulation" — The Robot Report · AI & Technology
LiberAI, a robotics-AI company, announced Liber-0, a "" trained on more than 2.2 million hours of human experience/video rather than robot-specific demonstrations. The goal is : a robot can handle objects or tasks it was never explicitly trained on. The company reports a 36% zero-shot success rate across a 60-task, 600-trial benchmark, positioning Liber-0 as a step toward general-purpose robot control learned from watching people.
This is a vendor announcement, distributed via press release and repeated by The Robot Report and other trade outlets with no independent testing cited. The benchmark numbers are LiberAI's own claims, so the story is solid as a product-announcement report but unverified as an independent performance result.
Genuinely new: the announcement and coverage are dated November 4, 2026, with no recirculation year flag.
We covered two June 2026 research stories on pretraining robot manipulation from human-centric data—action-effect memory pretraining and scalable vision-language-action learning from human videos—which are the same research direction Liber-0 sits in, though neither involved LiberAI or Liber-0 itself.