vizmaya desk · 4 November 2026

AIRoA-0: a robot taught by 16,151 hours of human hands

Its team collected 0 hours of robot data. On ten tasks it scored 78%. The best rival scored 27%.

How do you teach a robot with no robot?

A Universal Robots UR5 arm with a two-finger gripper, standing on a flight case in an exhibition hall. — Photo: Cobot.jpg by GrowSkills Robotics / Wikimedia Commons, CC BY-SA 4.0. A UR5 arm, for scale. Not from the paper.

Source: AIRoA-0 report (arXiv:2611.03815), PDF p1 and p4

Robot data comes from robots.

Most robot-learning models train on recordings made on robots.

So the pile of data grows with the number of robots. Robots are few and costly.

What if the pile grew with the number of people?

Source: PDF p14 (related work) and p1

A person's video shows no robot at work.

A recording made by a person never shows a robot acting, failing or recovering. And the robot has no head camera. It sees through two wrist cameras, each a small slice of the room.

0

robots in the recordings

2

cameras, both on the wrists

So what tool could a person and a robot share?

Source: PDF p1 (introduction)

Give the person the robot's hand.

The team built a gripper called YUBI. Close your fingers and the jaws close. A camera sits on the wrist.

Then the same gripper is bolted onto a robot arm. Same camera, same spot.

How much can people record with it?

Source: PDF p6 (3.1). Diagram is a schematic, not the paper's figure.

The pile is big, and growing.

16,151hours of demos

That is 10.65 million clips of about 5.5 seconds each.

An early workshop version listed 2,730 hours. On 8 June 2026 the team's paper listed 8,434. By 2 November it was 16,151. That is 1.91 times (computed), though the two counts may be defined differently.

How does a model learn from all that video?

Source: PDF p7 (Table 2); arXiv:2606.10244; workshop page yubi-airoa-io.pages.dev

Every clip has two labels.

One names the whole job. One names the step happening right now.

At run time the robot gets only the job label. The step labels are for training.

What should it look at while it watches?

JOB LABELReplace the worn parts of a robot vacuum

STEP LABELTake off the first used mop pad

STEP LABELTake out the brush

STEP LABELPut the dustbin back

Source: PDF p7 (3.2). Labels paraphrased from the paper's example.

It learns where to look.

A wrist camera sees far more than a task needs. The model learns to give weight to the grippers and the objects, and less to the table and walls.

Nobody drew boxes to teach it.

Does it work in a room it has never seen?

Source: PDF p8 to 9 (4.4, Fig 8). Schematic, not the paper's figure.

78% vs 27%.

Ten two-handed tasks, at a site where none of the training data was recorded.

Four rival models got the same task demos and their best setting per task. A-0 used one recipe for all ten. Average success: 78 for A-0, then 27, 23, 16 and 14.

What does a win look like up close?

Source: PDF p3 to 4 (Table 1). 15 trials per task, partial credit. Rivals: π0.5, XR-1, GR00T N1.7, MolmoAct2.

Two minutes. One instruction.

In one run, A-0 takes off two used mop pads, presses on two new ones and pulls out the brush. The last step lands at 106 seconds.

Across 15 trials this task averaged 73%.

What if someone shoves the table mid-task?

Source: PDF p4 (Fig 3, Table 1). One run shown.

Someone pushes the bins. The arm adjusts.

A green part is on its way to the green bin. A person pushes the bins aside, and the red bin ends up under the gripper. The arm moves over and drops the part in the green bin, 4.8 seconds after the start of the clip.

The paper shows this once. We treat it as a demo, not a statistic.

Does it only work on one robot?

Source: PDF p5 (2.2, Fig 4). One run shown.

Same weights. Four robots.

A bimanual Franka, the humanoid TX-G2, OpenArm and a pair of UR arms all ran the same weights. Nothing was retrained for any robot. Only the Franka is scored. The rest are shown in frames.

10

Franka: all ten tasks, scored

4

TX-G2 tasks shown

3

OpenArm tasks shown

2

UR tasks shown

Does more data keep helping?

Source: PDF p5 to 6 (2.3, Fig 5)

Each doubling of data cuts error by 2.4%.

The team trained one model on 6.25%, 12.5%, 25%, 50% and 100% of the data. More data meant lower error on held-out clips, even after the same post-training.

This measures prediction error, not robot success. It is five points over roughly 1,000 to 16,000 hours.

So what do you do with this?

Source: PDF p10 (Fig 9, 10). Values on the line are computed from the paper's fit L = 0.063 D^-0.035 (R² = 0.98).

Three checks for any robot demo.

A good robot video can hide a lot. Next time you see one, look for these three facts. Here is how A-0's paper answers them.

TEST SITEA site where none of the training data was recorded

NUMBER OF TRIALS15 per task, with partial credit for half-done tasks

WHO RAN THE RIVALSThe authors' own team

Who had what going in?

Source: PDF p2 to 3, p15

Contested: how to count "no robot data"

Who had what going in.

The team collected no robot data. But A-0 starts from public π0.5 weights, and π0.5 was trained with robot data.

	A-0	Each rival
Task demos	Same for all five	Same for all five
Handheld pre-training	16,151 hours	None. Public checkpoint.
Settings	One recipe	Best of many, per task

So the table compares whole systems. Even a perfect picker of the best rival on each task averages 35.9%, against 77.6% for A-0 (computed).

Where does A-0 break?

Source: PDF p3, p4, p8, p14; π0.5 paper arXiv:2504.16054. 35.9 and 77.6 computed from Table 1.

Where it breaks.

Earbuds: 30% after 25.8 hours of demos, the most of any task.

Cup pyramid: 80% vs 71%, the closest win. A-0 still tops every task, by 9 to 60 points (computed).

The paper has no limitations section. No outside team has repeated it, and the dataset was listed as "coming soon" on 4 November.

Who made this, and where can you check it?

Source: PDF p4 (Table 1), p15. Best rival per task and gaps computed.

Credit and sources.

Research by Yueh-Hua (Kris) Wu, Kei Ota and 18 co-authors at AIRoA (AI Robot Association). Funded in part by NEDO project JPNP25015. Photo credits sit under each photo and again below.

AIRoA-0: Long-Horizon, Precise, Cross-Embodiment Manipulation from Handheld Data (report, November 2026)
AIRoA-0 project page
YUBI: Yielding Universal Bidigital Interface (arXiv:2606.10244)
π0.5: a Vision-Language-Action Model with Open-World Generalization (arXiv:2504.16054)
OpenAlex record check (4 November 2026: no entry for A-0 yet)

Method: every number was checked against the report. Figures marked "computed" are our own arithmetic. We read all 18 pages. We did not watch the videos.

Cobot.jpg: GrowSkills Robotics, CC BY-SA 4.0
ER-FLEX mobile manipulator: Auledas, CC BY 4.0
Franka Emika1.jpg: Ims, CC BY-SA 4.0
UMAX U-Smart Robot Vacuum: Jirka Matousek, CC BY 2.0
R-24 - Universal Robots.jpg: Toxophilus, CC BY-SA 4.0
Zusammengelegte Handtücher.jpg: User:Mattes, public domain
Jabra True Wireless Earbuds Lineup: ScribblingGeek, CC BY-SA 4.0
All photos via Wikimedia Commons

vizmaya desk · 4 November 2026