LiberAI unveils Liber-0, a world action model scoring 36.0% on 60 GPT-6 tasks

Liber-0 scores 36.0% across 600 trials

LiberAI, a Beijing-based startup, has introduced Liber-0, which it calls a world action model. The model was pretrained on more than 2.2 million hours of human experience, combining human video, glove-recorded hand motion and embodiment alignment, the company said.

LiberAI evaluated Liber-0 on 60 tasks proposed by GPT-6, a suite separate from its internal scaling benchmarks. The suite had 20 pick-and-place tasks, 20 non-pick-and-place manipulation tasks and 20 challenging non-pick-and-place and long-horizon tasks. Liber-0 scored 63.5% on pick-and-place, 34.0% on non-pick-and-place manipulation and 10.5% on the challenging non-pick-and-place and long-horizon group. Its overall success rate was 36.0% across 600 trials.

The model did best on pick-and-place and weakest on the group that combined harder manipulation with longer sequences. LiberAI reported the results without robot demonstrations collected specifically for those target tasks.

What zero-shot means in this test

LiberAI defines zero-shot narrowly here. The evaluated checkpoint was not trained on robot demonstrations collected specifically for these target tasks, the company said. Pretraining still included robot data. That definition still allows robot data in pretraining.

The examples include placing a pen in a pencil case, pushing a drawer closed, and a longer sequence in which a pen goes into a pouch and the pouch goes into a drawer.

Why LiberAI uses human experience

Robot manipulation has relied on demonstrations recorded on a specific robot for a specific task, usually by a person teleoperating the machine. That data matches deployment closely, but collecting it can be slow and expensive when new tasks or robot platforms need more demonstrations, according to LiberAI. Human activity is far more abundant, and much of it is already on video.

Human video does not translate directly into robot commands. Cameras sit in different places, hand poses estimated from video are noisy, and human bodies move differently from robots, LiberAI said.

Liber-0 connects human experience with robot action in three ways. LiberSense Glove, LiberAI's data-collection system, records hand movements directly, helping link video to the motion behind it. LiberAI's robot platform uses camera views plus hand, wrist and waist movements that more closely match human motion. A Unified Embodiment Representation gives the model body shape and joint information so it can adapt an action to the body performing it.

The model predicts future video frames and actions together. That lets it learn from ordinary video without action labels, alongside glove and robot recordings that include measured motion.

Glove data fills gaps in robot demonstrations

Glove data can add behaviors missing from a robot policy's training data. LiberAI trained policies whose robot demonstrations covered only pick-and-place, while glove demonstrations included folding, pressing and opening. Performance was measured as a progress score from 0 to 100, which credits partial completion, across four pick-and-place and six non-pick-and-place tasks.

With robot data alone, the non-pick-and-place progress score was 2.2. Adding glove recordings of those actions raised it to 47.5. Adding glove recordings that also included pick-and-place raised it to 54.6, close to the 56.3 scored by a policy trained on robot demonstrations of every action. LiberAI said this controlled experiment is separate from the 60-task suite: the actions appeared in the glove demonstrations but not in the policies' robot demonstrations.

Glove data can cut robot demonstrations

A separate experiment tested whether glove recordings can reduce the need for robot demonstrations. In an eight-task evaluation on LiberAI's robot platform, the full robot and glove pools contained equal numbers of demonstrations. A policy trained on 25% of the robot pool plus 100% of the glove pool reached a progress score of 75.6. A policy trained on 100% of the robot pool scored 72.8.

In that setting, glove recordings delivered comparable performance with 75% fewer robot demonstrations. The saving applies to robot demonstrations specifically, because the mixed policy still used the full glove pool. LiberAI said glove collection does not require a robot at each site, so experience can be gathered from more people, in more places and potentially at lower cost.

Internal benchmarks and the open question

On LiberAI's internal benchmarks, the pretrained checkpoint scored 81.4 on pick-and-place and 72.3 on non-pick-and-place tasks when evaluated directly, according to LiberAI. Those internal numbers are higher than the 60-task suite results, which LiberAI described as a separate test of generalization.

The open question is whether the 36.0% result holds outside a 60-task suite proposed by GPT-6. The result is an evaluation on a 60-task suite. LiberAI's claim rests on transferring human experience to robot action.