Liber-0, a world action model from Beijing-based startup LiberAI, completed 36.0% of attempts across a 60-task evaluation without robot demonstrations collected for those tasks. The model was pretrained on more than 2.2 million hours of human experience. LiberAI ran 600 trials in total.
The 60 tasks were proposed by GPT-6 and sit apart from LiberAI's internal scaling benchmarks. They split into 20 pick-and-place tasks, 20 non-pick-and-place manipulation tasks, and 20 harder non-pick-and-place and long-horizon tasks. Examples include placing a pen in a pencil case, pushing a drawer closed, and a sequence in which a pen goes into a pouch and the pouch goes into a drawer.
Results by task type
Success fell as tasks became more complex. Liber-0 reached 63.5% on pick-and-place, 34.0% on other manipulation, and 10.5% on the hardest group of non-pick-and-place and long-horizon tasks.
LiberAI defines zero-shot narrowly. The evaluated checkpoint was not trained on robot demonstrations collected specifically for the target tasks. Its pretraining still included robot data.
On LiberAI's internal benchmarks, the pretrained checkpoint scored 81.4 on pick-and-place and 72.3 on non-pick-and-place tasks, measured as progress scores. Robot post-training added 2.2 and 1.5 points, for 83.6 and 73.8. LiberAI reads the small gain as a sign that pretraining already captures most of the capability those tasks need.
How human experience becomes robot action
Human video does not map cleanly onto robot commands. Cameras sit in different places, hand poses estimated from video are noisy, and human bodies move differently from robots. LiberAI says Liber-0 addresses this in three ways.
The first is LiberSense Glove, the company's data-collection system, which records hand movements directly and links video to the motion behind it. The second is a LiberAI robot platform whose camera views and hand, wrist and waist movements are built to resemble a human's more closely. The third is a Unified Embodiment Representation, which gives the model information about body shape and joint movement so it can adapt actions to the body that performs them.
Liber-0 predicts future video frames and actions jointly. That lets it train on ordinary video with no action labels, alongside glove and robot recordings that carry measured motion.
Glove data and the robot-data budget
LiberAI ran two controlled experiments that are separate from the 60-task suite. In the first, policies were trained on robot demonstrations covering only pick-and-place, while glove demonstrations included folding, pressing and opening. Progress scores ran from 0 to 100 and credited partial completion, across four pick-and-place tasks and six other 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. Including glove pick-and-place data as well brought it to 54.6, close to the 56.3 from a policy trained on robot demonstrations of every action.
The second experiment tested whether glove data can replace robot demonstrations. On an eight-task evaluation using LiberAI's robot platform, the full robot and glove pools held equal numbers of demonstrations. A policy trained on 25% of the robot pool plus all of the glove pool scored 75.6. A policy trained on the whole robot pool scored 72.8. That is comparable performance with 75% fewer robot demonstrations, although the mixed policy still used the full glove pool.
LiberAI also built an interactive simulation where users upload their own objects and watch Liber-0 manipulate them. That version was post-trained on 100 hours of simulated teleoperation data and received no object-specific training.
Songming Liu, founder and CEO of LiberAI, said human video shows what people do but not exactly how their hands move or how that movement fits a different body. Grounding the experience in measured motion and aligning it with the robot's body, he said, is what turns it into robot capability.
LiberAI lists limits of its own. The glove records hand poses and trajectories but not contact or force, which matter for insertion, fastening and handling soft materials. Tactile and force sensing, and data collection across a wider range of homes, remain future work.
What the results do and do not show
The strongest evidence in this release is the demonstration-count result. Teleoperated robot data is slow and costly to collect, and it must be repeated for new tasks and platforms. A glove can be worn without a robot on site. If the 75% reduction holds beyond eight tasks, the cost of building a manipulation dataset would shift toward cheaper human-side collection. The result comes from one company's platform and one task set, so it is a signal rather than a settled figure.
The headline 36.0% needs careful reading. The average hides a steep gradient, and the hardest group of multi-step tasks succeeded 10.5% of the time. A model that fails roughly nine in ten long-horizon attempts is not close to deployable for those jobs. The 63.5% on pick-and-place is the more useful guide to where the approach works now, and that task family is also the one most existing robot datasets already cover.
Comparison with other systems is hard on the evidence given. The results are reported by LiberAI, and the task suite was generated by a language model rather than drawn from a shared benchmark. The source gives no results for competing models on the same 60 tasks. Success rates, used for the zero-shot suite, and progress scores, used for the controlled experiments, are different metrics and should not be read against each other.
The zero-shot label also carries caveats. Pretraining included robot data, so the claim is about the absence of task-specific demonstrations, not the absence of robot experience. How much of the 36.0% comes from the human-experience pretraining, rather than from that robot data, is not isolated in the figures released.
Three tests would settle the open questions: independent replication on a shared task set, results on contact-rich tasks once force sensing is added, and evidence that the glove advantage holds on robot bodies other than LiberAI's own. The design of the world action model, which learns from unlabeled video and measured motion together, is plausible. Its value now depends on whether those later tests confirm it.
