The exam the student helped write
Start with the 36.0%. It is a real number from a real set of trials: 600 of them, 10 per task. But every part of the test was arranged by the company reporting the score. LiberAI assembled the 60-task suite. GPT-6 proposed the tasks, which makes the suite sound independent until you ask who selected, filtered and set them up in a lab. No other model has published a score on this suite, and Liber-0 was not run on a standard external benchmark in the release. The release provides no baseline from another model on this suite, so 36.0% cannot establish how Liber-0 compares with competing systems. Until another model is evaluated under the same conditions, the figure is a company-reported starting point, not a comparative result.
The breakdown is more informative than the average. Pick-and-place, the most rehearsed skill in robot learning, scored 63.5%. Other manipulation scored 34.0%. The long-horizon group scored 10.5%, roughly one successful trial in ten. These tasks involve more extended sequences of actions, although the release does not establish how representative they are of real-world work. LiberAI's own wording concedes performance there "remains limited."
Then there is the phrase doing the most work: zero-shot. LiberAI defines it precisely, to its credit. The evaluated model "was trained on robot data during pretraining," but not on robot demonstrations collected for these 60 tasks. That is a meaningful research claim. It is also narrower than the homepage version, "60 Tasks It Was Never Trained On," which a casual reader will hear as a robot walking up to wholly new work cold.
The 2.2 million hours, itemized
The launch leads with scale: more than 2.2 million hours of human experience. Read further into the same release and the company itemizes that scale in a way that undercuts it. In controlled single-source pretraining, followed by the same robot post-training, 10,000 hours of glove recordings outperformed 2 million hours of web human video on both pick-and-place and non-pick-and-place benchmarks. The result suggests that data quality and representation matter more than raw volume in this experiment. Glove recordings contain measured hand poses and trajectories, while web video primarily supplies visual observations. LiberAI's comparison does not establish that this difference alone explains the performance gap, but it makes the case that instrumented human-action data can be substantially more useful than a much larger volume of ordinary video.
The supporting experiments point the same way. In a separate experiment, LiberAI measured partial task-completion progress across eight tasks. A policy trained with 25% of the robot demonstration pool plus the full glove pool scored 75.6, compared with 72.8 for the full robot pool alone. The company says the two complete pools contained equal numbers of demonstrations, making this a comparison of data composition rather than simply adding more demonstrations. When robot demonstrations covered only pick-and-place, the policy's progress on other manipulation was 2.2; adding glove recordings of those missing actions raised it to 47.5, and broader glove data to 54.6, near the 56.3 of a reference trained on robot demonstrations for both groups. Note the release's own caution here: these are progress scores, partial-completion marks out of 100, kept separate from the 36.0% success figure. The company is careful with its numbers. Its headline is less careful than its experiments.
What would settle it
Three things would turn this from a company-reported launch into a checkable result: a run on a public benchmark that other robot models also report; a rival score, or an independent replication, on the 60-task suite; and the basics the release omits, starting with model size and how anyone outside LiberAI can access the model or the simulation. The technical material describes an interactive simulation adapted with 100 hours of simulated teleoperation, which is a start, but it is still LiberAI grading LiberAI. The underlying idea, that measured human action transfers better than raw video, is supported by the company's own controlled comparisons and is worth taking seriously. The 2.2 million hours should be understood as the largest ingredient by volume, not necessarily the most valuable one. The company's own comparison suggests that smaller quantities of instrumented human-action data contributed more to performance in the experiments it reported.
Sources
- [1] LiberAI, "LiberAI Introduces Liber-0 for Zero-Shot Robot Manipulation", November 4, 2026, EIN Presswire (primary; full text read first-hand at Gate 1)Read source
- [2] LiberAI, "LiberAI Introduces Liber-0 for Zero-Shot Robot Manipulation", November 3, 2026, ACCESS Newswire version via CRWE World (same release, earlier wire date; full text read first-hand at Gate 1)Read source
- [3] LiberAI official site, homepage claims ("60 Tasks It Was Never Trained On"; 36.0% zero-shot success), read at Gate 1Read source
- [4] Robotics and Automation News, "LiberAI launches Liber-0 world-action model for general-purpose robots", November 4, 2026 (trade summary of LiberAI's technical material; source of the December 2025 founding date, the eight-task and cross-platform detail, and the interactive-simulation note)Read source