LiberAI's Liber-0 robot learns from human video and glove data

LiberAI's Liber-0 model has demonstrated scalable general-purpose robot learning from human video and glove data. The World Action Model achieved a 36.0 percent overall success rate across 600 trials on 60 distinct tasks, all without requiring a single robot demonstration collected specifically for those tasks.

This breakthrough addresses a critical bottleneck in manufacturing and robotics: the need for scalable training data. Liber-0 is trained on more than 2.2 million hours of human experience. It learns by jointly predicting future video frames and actions, gleaning skills from ordinary videos that lack explicit action labels. The model is designed not as a catalog of specific programs but as a foundational system for learning the relationships between visual observations, physical actions, and outcomes.

Task-specific performance breakdown

Success rates varied significantly depending on the type of manipulation required. The model performed strongest on straightforward pick-and-place operations. Performance dropped on more complex non-pick-and-place manipulations and fell further on challenging, long-horizon tasks.

Each of the 60 tasks, which were proposed by GPT-6 and distinct from the company's internal benchmarks, was tested ten times. LiberAI's own internal benchmarks showed a pretrained checkpoint achieving an 81.4 progress score on pick-and-place tasks and 72.3 on non-pick-and-place tasks.

Human video and glove data enable scalable training

Liber-0 bypasses traditional robot training, which relies on hardware access and operator effort to collect robot demonstrations. Instead, it leverages a scalable alternative: human experience. The model combines large-scale human video with hand poses and trajectories recorded directly by the LiberSense Glove. This connects visual observations with measured physical actions.

Glove data collection does not require a robot to be present, making it a vastly more scalable source of training data for physical AI. Key technical challenges in transferring human experience to robots include differences in visual viewpoint, gaps in action supervision, and embodiment gaps between human and robot bodies. Liber-0 uses a Unified Embodiment Representation that incorporates human body characteristics and robot arm capabilities, alongside visual and embodiment alignment techniques, to bridge these gaps.

Performance advantages over robot-only training

Using human glove data provides a clear performance benefit over relying solely on robot-collected demonstrations. In an eight-task evaluation, a policy trained on just 25 percent of the available robot demonstration data, combined with 100 percent of the glove demonstration pool, achieved a higher progress score than a policy using all the robot data alone.

The glove-augmented policy scored 75.6, compared to 72.8 for the policy trained on 100 percent robot demos. The value of glove data is particularly pronounced for non-pick-and-place actions. In tests, adding glove data covering these actions increased progress scores dramatically, from a baseline of 2.2 up to 47.5 and 54.6, approaching the performance of a reference set using full teleoperation.

On specific tasks using the LEO robotic platform, which incorporates dexterous hands and camera configurations aligned with human data, performance with matched glove data was strong. The LEO robot achieved scores of 91.3 on knob turning, 80.0 on arranging slippers, and 100.0 on opening a drawer. For comparison, the Tianji robot scored 76.3, 35.0, and 75.0 on the same tasks. Liber-0 can work with a higher-level AI agent to break down objectives into intermediate goals for physical action translation. The model was post-trained on 100 hours of simulated teleoperation data to improve object-level generalization. Its recent evaluation in New York City represents a concrete step toward integrating such general-purpose systems into real-world home environments.