LiberAI unveils world action model Liber-0, trained on more than 2.2 million hours of human experience

LiberAI unveiled Liber-0 on Nov. 3, 2026, a world action model that used more than 2.2 million hours of human experience for pretraining. In an evaluation presented by the company, the model attempted 60 robot manipulation tasks proposed by GPT-6, 10 times each, for a total of 600 trials, and achieved an overall success rate of 36.0%. The company said it performed the tasks without using robot demonstration data collected specifically for them.

In LiberAI’s evaluation, the model attempted 60 manipulation tasks without task-specific robot demonstrations, but it did use robot data during pretraining. The key question is how far an approach that transfers human experience to robot actions can make up for collecting demonstrations for each task. The success rate on the long-horizon task group was 10.5%, however, indicating that challenges remain in the model’s range of applicability.

The central point of the announcement is not just the collection of large amounts of human video, but the use of data measuring hand movements and robot embodiment in the same learning pipeline. Human video shows how objects are handled, but does not directly record the joint movements a robot must execute. LiberAI is seeking to bridge this gap by using the poses and trajectories recorded by the LiberSense Glove as cues for movement, and combining them with LEO’s embodiment configuration and joint information. This can be understood not as having a robot simply imitate human experience, but as a design for mapping that experience to actions the robot can execute. In the company’s comparison experiments, under conditions in which pretraining used a single data source and the same robot training followed, 10,000 hours of glove-recorded data outperformed 2 million hours of human video from the web on both pick-and-place and other manipulation benchmarks. This suggests that the sheer volume of data does not determine its value for learning, and that the correspondence between data and movement may affect how useful it is. The comparison does not establish that the same results would hold for other robots or environments, however, so the scope of the evaluation should be considered. The overall success rate in the 60-task evaluation was 36.0%, while the long-horizon task group scored 10.5%. A gap remains between simpler object-moving tasks, which had higher success rates, and operations requiring multiple steps. This difference matters in assessing which types of work can benefit from using human experience. We previously reported that LiberAI took the top spot on RoboDojo, but the benchmark ranking and the manipulation success rate in this announcement are separate evaluations and cannot be compared directly. The industrial implications include the possibility that training data need not depend solely on robot demonstrations, and that the measurement and embodiment information required for this approach must be developed alongside the data. Expanding the volume of human video may not be enough to enable execution if movement labels are lacking, while adding measured data may still have limited transferability if differences between robot bodies cannot be handled. The next questions are whether the announced success rate can be reproduced on different robot hardware and in real-world environments, how performance on long-horizon tasks changes, and how much the company discloses about the composition of its training data and model size. The announcement did not provide figures for model size or training compute.

The success rates were 63.5% on 20 pick-and-place tasks, 34.0% on 20 other manipulation tasks, and 10.5% on 20 difficult long-horizon tasks excluding pick-and-place. The company describes this as zero-shot generalization at the task level. However, the model used robot data during pretraining; the claim means that it did not use robot demonstrations collected specifically for the evaluation tasks. Liber-0 links human videos with hand poses and trajectories recorded by the company’s LiberSense Glove data-collection system, and jointly predicts future video and movement. The company’s LEO robot platform features dexterous hands, articulated wrists and a waist, and a camera configuration designed to align with human data collection. The company also says it uses a “Unified Embodiment Representation” to convey body dimensions and joint movements.