LiberAI Reports 36% Success in Liber-0's 600-Trial Robot Test

Hero image: archival Galaxea R1 Pro and Unitree G1 research platforms from the separate EgoScale study, Ruijie Zheng et al. (2026), Figure 9, reproduced under CC BY 4.0. Neither robot is LiberAI's LEO; the figure does not depict LiberSense equipment or Liber-0 tests. Original image reproduced for research context; display sizing and card crops may vary.

How Liber-0 approaches unfamiliar work

Learning terminology can obscure the boundary being tested. A new object, a new room and a new task are different challenges. A machine might recognize a behavior from human examples while never having practiced that exact job through robot demonstrations. Conversely, familiar movements can become difficult when the camera view changes. Readers should ask which boundary an experiment crosses before interpreting its headline as general autonomy.

This is a commercially relevant question even before robots become broadly autonomous. Every new task that needs dedicated demonstration collection adds preparation before a machine can do useful work. A transferable starting point could shorten that preparation. The present results measure attempts at unfamiliar tasks, however, rather than the time or money saved in a customer installation.

Under the hood: alignment before scale

Human video offers abundant examples of object handling, but it often omits the information a robot controller needs. A hand can conceal the contact point. Camera motion can obscure the object's trajectory. A human wrist can reach a pose that a different mechanism cannot reproduce. More recordings do not automatically remove those mismatches.

The earlier, separately published EgoScale research offers a useful explanation of action supervision. Its training targets include relative wrist motion and hand-joint actions, giving the model movement information rather than only pictures. The researchers then use aligned human and robot examples with similar tasks, scenes and viewpoints. This intermediate training connects a representation learned from human activity with the robot's own sensing and control system.

What the 600-trial test actually measures

LiberAI reports the following company-run evaluation. Each category contains 20 tasks, attempted 10 times each.

A repeated attempt is not another kind of work. Testing one task several times is valuable for measuring consistency, but it should not be confused with testing the same number of distinct tasks. The distinction also affects what failure analysis can reveal: recurring errors on one task may point to a specific interaction rather than a general inability to follow instructions.

Equal weighting also makes the overall result sensitive to the chosen mix. A suite dominated by easy placement jobs would produce a different headline even if the model itself were unchanged. Keeping the three group results visible is therefore essential when comparing future versions: a higher average should reveal whether the system improved difficult manipulation or simply faced a more favorable selection of tasks.

The category breakdown is more informative than the average for an operator. A workflow's reliability depends on its hardest required interaction and whether the system can recover when something goes wrong. A successful grasp followed by a failed closing step can still leave the job unfinished. The large difference between simple relocation and multistep completion makes recovery and sequencing central questions for the next evaluation.

From laboratory result to operating specification

A manipulation evaluation cannot, by itself, answer how a machine will operate around people. A deployment specification needs to say which objects and environments are supported, when a person must intervene and what happens when a task cannot be completed safely. Those are engineering requirements beyond demonstrating an individual skill, even when the skill is learned rather than explicitly programmed.

For a prospective integrator, the practical distinction is between a capable starting model and a dependable installation. The latter needs operating limits, recovery procedures and measured intervention rates over sustained use. A robot that detects its own failure and requests help can be more manageable than one that repeatedly continues with an incorrect assumption. Completion statistics should therefore be accompanied by information about how errors become visible.

What this adds to human-to-robot learning

LiberAI enters an established research direction. EgoScale describes 20,854 hours of action-labeled egocentric video, followed by aligned human-and-robot mid-training and task fine-tuning. DexUMI tackles the transfer problem with a wearable exoskeleton and image processing that substitutes robot-hand appearance. Both address the information lost when ordinary human footage is treated as an executable robot demonstration.

DexUMI makes a particularly concrete design choice: its exoskeleton adapts human motion to movements feasible for a robot hand, while its software replaces the human hand's appearance in training imagery with a robot hand. That separates two obstacles often grouped together as the embodiment gap. The hand must be able to execute the movement, and the controller must recognize the relevant scene when the visible hand changes.

EgoScale also evaluates transfer to the Unitree G1 with a three-finger hand, distinct from the higher-degree-of-freedom hands used elsewhere in its research. That is useful context for interpreting embodiment: a human movement representation need not map to identical hardware to be useful. But success on another body remains an empirical question, requiring testing rather than an assumption based on a model label.

There is also a trade-off between matching one robot closely and serving many different robots. Closely aligned cameras and joints can simplify transfer, but a method that works well in that arrangement may need further adaptation on another body. Testing the same learned behavior with different hands or viewpoints would help identify how much capability belongs to the shared model and how much depends on the carefully matched platform.

The research cannot be ranked by headline percentages alone. Different objects, task definitions, training stages and success rules change what a score means. Nor does the largest video-hour count necessarily provide the best action supervision. For this field, the meaningful scaling question is how much usable physical knowledge survives the journey from recording to robot movement.

What evidence should come next

Independent repetition of a fixed task suite would help separate the method's contribution from setup choices. Results broken down by task, failure type and recovery behavior would be especially useful: an object slipping from a hand requires a different improvement from a planner choosing the wrong next step. Reporting completion time alongside success would show whether better reasoning introduces a throughput penalty.

The economic evidence needs its own measurement. Reducing robot demonstrations can shift work toward wearable collection, annotation and data processing rather than eliminating it. A useful comparison would count the total preparation effort needed to reach a fixed reliability target, including failed collection attempts and later corrections. That would connect a learning result to the cost an operator actually faces when introducing a new task.

A longer evaluation could then test new objects, changed camera views and repeated operation without carefully restored starting conditions. Those changes would probe whether alignment remains useful outside the collection setup. They are proposed tests of the announcement's implications, not milestones LiberAI has committed to deliver.

Frequently Asked Questions

A clearer measure of the transfer problem

The important question after Liber-0's introduction is whether transferable learning can shorten preparation while improving complete-workflow reliability. Human-data research gives the field several ways to attack the problem, from better movement labels to closer visual alignment. The decisive evidence will connect those choices to repeatable task completion under clearly specified conditions.