Liber-0 Hits 36% Zero-Shot Robot Success

Liber-0 shows that human action data can transfer useful physical intuition to robots, but the steep drop on long-horizon tasks makes clear that zero-shot manipulation remains far from reliable autonomy.

• –The model was pretrained on more than 2.2 million hours of human, glove, robot, and video experience, then evaluated on tasks absent from its robot demonstrations. ([LiberAI](https://www.liberai.com/blog/liber0/))
• –Its strongest result is relatively structured pick-and-place; multi-step contact-rich manipulation remains the major bottleneck.
• –LiberAI reports a large advantage over GPT-6 Astra alone, which achieved only 8.3% on pick-and-place and 1.7% on difficult multi-step tasks. ([LiberAI](https://www.liberai.com/blog/liber0/))
• –The results are company-reported, with no independent replication, public weights, or API identified so far. ([ExplainX](https://explainx.ai/blog/liberai-liber-0-world-action-model-36-percent-zero-shot-60-tasks-glove-data-gpt-6-astra-5-percent-claimed-vs-verified-2026))