LiberAI Unveils Liber-0 World Action Model: Pretrained on 2.2M Hours of Human Experience, 36% Zero-Shot Success Across 60 Manipulation Tasks

Background and pain point

Robot manipulation has long depended on teleoperated demonstrations collected for a specific robot and task, which is expensive and hard to scale. Human video is abundant but faces three gaps: human and robot bodies differ, hand motion estimated from video is noisy, and camera viewpoints differ. Liber-0, released by LiberAI on November 4, aims to turn large-scale human experience directly into executable robot skills.

What is new

Liber-0 is a World Action Model built on a Mixture-of-Transformers architecture in which language, vision and action experts jointly predict future frames and actions. A Unified Embodiment Representation encodes human and robot morphology and kinematics in one space. On the data side, the 0.3 mm, 131 g LiberSense Glove records precise finger motion alongside first-person and wrist cameras; on the hardware side, the in-house LEO robot has three-degree-of-freedom wrists and an articulated waist, with camera views aligned to the human recordings. Pretraining runs in three stages: general visual prediction, joint world-action learning, then manipulation-focused training with more glove and teleoperation data.

Numbers

All figures are self-reported on the official blog. On 60 zero-shot tasks proposed by GPT-6 over 600 trials, overall success was 36.0% and multi-step success 10.5%; GPT-6 Astra alone as a policy scored 8.3%, 5.0% and 1.7% on the same task groups. On internal progress benchmarks, the full recipe after robot post-training scored 83.6 on pick-and-place and 73.8 on non-pick-and-place. Raising demonstrators from 1 to 100 at a fixed data volume lifted progress from 60.6 to 69.5.

Getting started

So far there is only a technical blog and demo videos; no open weights, API or pricing has been announced, so developers cannot call the model yet. The blog says Liber-0 can pair with a reasoning agent such as GPT-6 Astra to plan intermediate goals for multi-step tasks, and offers an online simulation app where users can upload custom objects. For embodied-AI teams the main takeaway is the data approach: scaling human demonstrations with low-cost gloves to reduce reliance on robot teleoperation data.