Chinese robotics startup LiberAI has launched Liber-0, a world action model for robot manipulation. The company says that across 60 tasks and 600 trials, Liber-0 reached 36 percent overall success "without robot demonstrations collected specifically for those tasks." The claim appears in LiberAI's launch post on X and on its website.
The numbers are self-reported. The 60 tasks were proposed by GPT-6 and run by LiberAI. This post lays out what the company claims, what trade-press coverage confirms, and what you should check before treating the results as settled.
TL;DR: claims and checks
| Question | Answer |
|---|---|
| What is Liber-0? | A World Action Model that predicts future video and actions jointly |
| Training data? | More than 2.2 million hours of human experience: video, glove recordings, robot data |
| Headline result? | 36.0% success over 600 trials on 60 tasks, zero-shot (company-reported) |
| By task group? | 63.5% pick-and-place, 10.5% complex multi-step (company-reported); the middle group's rate was not in the sources we read |
| GPT-6 Astra comparison? | Amplifying posts say 5% on the same tasks; run by LiberAI, not independent |
| Glove versus web video? | 10,000 glove hours beat 2M hours of web human video in matched runs, per the company |
| Independent replication? | None found on launch day |
| Open weights? | Not stated in what we read |
What is a world action model?
Most robot foundation models map camera images and a text instruction to actions. A world action model adds a prediction of what the world will look like next. LiberAI describes Liber-0 as learning relationships between visual observations, physical actions and what happens afterwards. According to coverage in Robotics & Automation News, training runs in stages:
- Broad visual prediction from videos of diverse human activity.
- Joint prediction of future observations and actions, combining human video with glove and robot recordings.
- A manipulation-focused stage with more weight on precise motion and sustained contact.
The point is to learn from human experience at a scale robot hardware cannot match. A person can fold a towel in a video. A robot needs a way to turn that into joint commands for its own body.
What is the glove idea?
The core bet is a data glove. LiberAI's LiberSense glove, reported at 131 grams, records hand poses and trajectories while people handle objects. A first-person camera and two wrist cameras record at the same time, positioned to resemble the views of LiberAI's own robot, LEO.
The advantage is cost. Collecting glove data needs no robot and no teleoperation rig. A person wears the glove and does ordinary work. The coverage also lists limits: the glove records hand pose and trajectory but not contact or force, which matter for inserting, fastening and handling soft materials.
LiberAI names three gaps it tries to close: a visual gap between human and robot cameras, an action-supervision gap because hand motion inferred from plain video is noisy, and an embodiment gap because human and robot bodies differ. Liber-0 uses what it calls a Unified Embodiment Representation, which feeds in body details such as height, arm length and hand geometry alongside robot arm and hand geometry.
What exactly was tested?
The headline test used 60 manipulation tasks "proposed by GPT-6" and 600 trials, so ten trials per task on average. The company reports:
| Task group | Reported success |
|---|---|
| Pick-and-place | 63.5% |
| Second group (other manipulation) | not stated in the sources we read |
| Long-horizon / complex multi-step | 10.5% |
| Overall | 36.0% |
The company's term is task-level zero-shot generalization: no robot demonstrations were collected for those tasks. That is narrower than it sounds. The model was still pretrained on human video, glove data and robot data from other tasks.
Why the GPT-6 Astra number needs care
Posts that amplified the launch say GPT-6 Astra completed 5 percent of the same tasks. We could not find a LiberAI methods page for that comparison in the text we read. Open questions include which harness drove the robot, how Astra was connected to the controller, what prompts it had, and whether it got the same number of trials. Compare it with our earlier coverage of GPT-6 Astra on a robot control benchmark, where harness choices moved results a great deal.
A language model driving a robot arm through an API is not the same kind of system as a model trained on motion data. A 36 percent versus 5 percent gap is more a comparison of approaches than a ranking of intelligence.
The glove-versus-video claim
The most interesting reported result is about data efficiency. In matched runs, 10,000 hours of glove recordings beat 2 million hours of web human video, according to the company's scaling claims as relayed in launch coverage. The company also reports that broader glove data raised one benchmark score to 54.6, near 56.3 for robot demonstrations covering both task groups.
If that holds, it is a significant data-collection finding: a smaller amount of action-grounded data can outweigh far more unlabeled video. It is also easy to overread. The comparison depends on the benchmark, the model, the video source and the way each dataset is filtered. Wait for the technical report and for outside groups to try the same ablation.
Human-to-robot transfer between two robots
Coverage of the technical material describes a test that trained separate policies for LEO and another platform, Tianji, using the same glove recordings and equal amounts of teleoperation data. The reported scores on three tasks were:
| Task | LEO | Tianji |
|---|---|---|
| Turning a knob | 91.3 | 76.3 |
| Arranging slippers | 80.0 | 35.0 |
| Opening a drawer | 100.0 | 75.0 |
| Average | 90.4 | 62.1 |
The gap shows that the same human data transfers unevenly. LiberAI designed the glove cameras to match LEO's views, so LEO has a home advantage. That is a fair design choice for a product, and a reminder that "human experience" is not robot-agnostic.
Demos, simulation and the agent layer
LiberAI's website points to an application workflow: adapting pretrained manipulation skills to a real task, shown as autonomous boba tea preparation by the LEO robot. A demo like this shows a fine-tuned pipeline working on one task. It does not tell you the zero-shot rate, because the model was adapted for it.
Coverage also describes an interactive simulation in which the policy received 100 hours of simulated teleoperation post-training. Users can then drop in their own objects and watch Liber-0 try to handle them without object-specific training. That is a useful way for outsiders to probe the model, and it is worth testing with odd shapes and transparent or deformable items, where manipulation models usually fail.
For longer jobs, the company says Liber-0 can work under a higher-level AI agent. The agent splits a big goal into intermediate steps, Liber-0 turns each step into motion, and new camera observations feed back. This split is common across robotics: a language model plans, a learned controller acts. The 10.5 percent long-horizon score suggests the hard part is still keeping errors from compounding across many steps.
How does this compare with other robot models?
The field is crowded with world-action and vision-language-action claims. See our coverage of Dream Labs' Dream-1 robot API, the uni-robot-agent result on RoboDojo, the RoboColosseum evaluation of GR00T and pi0.5 and Xiaomi Robotics-1's 100,000 hours of UMI data. Each makes a different bet on data: teleoperation, handheld grippers, glove capture or web video.
LiberAI has also published a Liber-0 Preview checkpoint for the RoboDojo benchmark on Hugging Face. RoboDojo is a different benchmark from the 60-task test, and we did not verify either one.
What the numbers do not tell you
- A 36 percent success rate is not reliable enough for a home. Failing roughly two of three unseen tasks, and nine of ten long multi-step ones, is early research.
- Tasks were chosen by the lab's pipeline. Even with GPT-6 proposing them, LiberAI selected and ran them.
- Ten trials per task is thin. Per-task rates carry wide error bars.
- Hardware matters. Results are for LiberAI's own robots and sensors.
- No force data yet. LiberAI says contact and force are future work.
How to evaluate a robot-model claim like this
- Find the task list. Ask for the 60 tasks and success criteria. Check for overlap with training scenes.
- Check trial counts and seeds. Look for per-task results, not just the average.
- Look for third-party runs. An outside lab on different hardware is worth more than any internal chart.
- Read the baseline setup. If a language model is the baseline, check how it was wired to the robot.
- Separate ablations from products. A data-efficiency ablation is a research result. A demo video is not a deployment.
Bottom line
Liber-0 is a notable idea: scale robot learning with cheap, robot-free glove data and large human video, and condition on the body that will act. The reported 36 percent zero-shot result and the glove-versus-video ablation are interesting, and they are the company's own. Until the technical report is open to scrutiny and an outside group reproduces it, treat the numbers as promising claims.
Details reflect LiberAI's website and X post and trade-press coverage as of November 4, 2026. Results and availability may change.