LiberAI's Liber-0: 10,000 hours of glove recordings beat 2 million hours of web video

Beijing startup LiberAI has released Liber-0, a robot model it calls a "world action model," pretrained on more than 2.2 million hours of human experience. On 60 tasks proposed by GPT-6, for which the model had seen no robot demonstrations, it succeeded in 36.0% of 600 trials, the company reported on Nov. 4.

LiberAI's own numbers credit something other than the size of that dataset. In its scaling tests, pretraining on 2 million hours of web video of people scored 64.6 on pick-and-place and 51.2 on other manipulation, both below what 10,000 hours recorded with its LiberSense data glove achieved. The glove is 0.3 mm thick, weighs 131 grams and measures finger motion directly, while a head camera and two wrist cameras film the task. A 200-fold gap in hours went the glove's way. That makes LiberAI's collection kit as important as its model, and puts it up against companies such as GenRobot, which just raised $100 million to sell recordings of people at work.

Glove recordings stood in for three-quarters of the robot demonstrations

With robot demonstrations cut to 25% of the pool and the full glove pool added, a policy reached a progress score of 75.6 across eight tasks, above the 72.8 from the full robot pool alone. The two pools held equal numbers of demonstrations, so total data rose; what fell was robot time, the part of teleoperation that needs a robot and an operator for every demonstration. The glove also taught actions the robot data lacked. When robot demonstrations covered only pick-and-place, other manipulation scored 2.2; adding glove recordings of those actions raised it to 47.5, and all the glove data to 54.6, close to the 56.3 of a policy trained on every teleoperated demonstration.

Multi-step tasks are where zero-shot falls off

The zero-shot result splits sharply: 63.5% on pick-and-place, 34.0% on short manipulation tasks and 10.5% on challenging multi-step ones. For long tasks, Liber-0 can be paired with a reasoning agent such as GPT-6 Astra to plan intermediate goals. GPT-6, not LiberAI, proposed the tasks, all of them on a tabletop. The hardware is built for the transfer too: LiberAI's LEO robot pairs dexterous hands with articulated wrists and an adjustable waist so that it can copy human movement, and on the same glove data it averaged 28.3 points higher than a Tianji platform with Marvin M6 arms.

LiberAI was founded in December 2025 by Liu Songming, a Tsinghua computer science graduate who led the RDT series of robot foundation models, and in June raised a Pre-A round of several hundred million yuan led by Shunwei Capital, its third round of the year, the STAR Market Daily reported.

Run as the robot's only policy, Astra managed 8.3% on pick-and-place, spent 87.3 seconds reasoning per decision, and took 17.8 minutes on average to finish a task it got right.