Evidence-first physical AI intelligence

Know what humanoid robots can actually do.

Most robot databases repeat specifications. We report what independent evidence supports — how capable a robot has been shown to be, how commercially real it is, and how much you should trust either number. The three are never merged.

Explore 21 robots Compare robots
21Robots tracked
4Meet the evidence publication standard
v1.2.0Frozen scoring methodology
2026-09-06Last catalogue update

Three answers, never one score

A single rating would hide the thing you need to know.

A well-evidenced prototype and a poorly-evidenced product are not the same, and one number cannot say both. So we publish the dimensions separately and show our working.

Capability evidence

What the robot has been shown to do, weighted by how independent and how durable the demonstration was. Manufacturer video earns less than an examined third-party test.

How evidence is tiered →

Commercial reality

Whether anyone is actually paying for it and running it. Reservations, pilots, contracts and verified production are kept structurally separate — a 2035 agreement is not a robot on a line today.

How commercial reality is scored →

Evidence confidence

How much to trust the two numbers above. Thin sourcing caps confidence no matter how impressive the underlying claims look. It is a statement about our evidence, not about the machine.

What caps confidence →

Selected profiles

Four robots, four different evidence positions.

Chosen to show the range of what the evidence supports — not a ranking, and not a recommendation. Every profile carries the full source ledger behind its numbers.

Branded identity panel for the Tiangong Ultra by Beijing Humanoid Robot Innovation Center. This is a graphic generated by Physical AI Radar, not a photograph of the robot.

Tiangong Ultra

Beijing Humanoid Robot Innovation Center · China
Evidence ready
Capability6.0
CommercialINSUFFICIENT DATA Insufficient data is not a score of zero and is not better or worse than one — this dimension could not be calculated at all.
Confidence 82.2Very High
Completeness 59.0%
Branded identity panel for the Walker S2 by UBTECH Robotics. This is a graphic generated by Physical AI Radar, not a photograph of the robot.

Walker S2

UBTECH Robotics · China
Evidence ready
Capability12.0
CommercialINSUFFICIENT DATA Insufficient data is not a score of zero and is not better or worse than one — this dimension could not be calculated at all.
Confidence 75.3High
Completeness 41.0%
Branded identity panel for the AGIBOT G2 by AGIBOT. This is a graphic generated by Physical AI Radar, not a photograph of the robot.

AGIBOT G2

AGIBOT · China
Evidence ready
Capability4.3
Commercial42.8
Confidence 54.5Moderate
Completeness 59.0%
Branded identity panel for the Unitree R1 by Unitree Robotics. This is a graphic generated by Physical AI Radar, not a photograph of the robot.

Unitree R1

Unitree Robotics · China
Evidence ready
Capability8.7
Commercial15.0
Confidence 53.6Moderate
Completeness 88.0%
See all 21 robots

Demo is not deployment

Where a robot actually sits on the commercial ladder.

These four positions are kept structurally separate in the data model. This is an explanation of the ladder, not a score, and no robot is placed on it here.

ReservationA deposit or an expression of intent. No contract, no delivery, no operation.
Test / POCThe robot is on site and being evaluated. Nobody is depending on it yet.
Contract / offtakeMoney has changed hands or been committed. Delivery may still be years out.
Verified productionIndependently confirmed units doing real work on a real line, today.

One claim, fully traced

Every number on this site opens into its source.

“Longcheer Technology (Apr 2026): AGIBOT G2 robots performed pre-shipping tablet testing on Nanchang production line. PoC acceptance confirmed early 2026.”

AGIBOT G2 Longcheer Technology · 2026-04-15 Independent Customer Independent Durably captured

Latest evidence changes

What changed, and why.

Drawn from recorded score history. Nothing on this page is generated or inferred.

DateRobotChangeEngine
2026-09-06 Booster T1 Batch B — Unitree R1 EVIDENCE READY via EdUHK Academic Scoring Methodology v1.2.0 (ENGINE v1)
2026-09-06 Booster T2 Batch B — Unitree R1 EVIDENCE READY via EdUHK Academic Scoring Methodology v1.2.0 (ENGINE v1)
2026-09-06 Walker S2 Batch B — Unitree R1 EVIDENCE READY via EdUHK Academic Scoring Methodology v1.2.0 (ENGINE v1)
2026-09-05 LimX Luna Batch C/D — Tiangong Ultra EVIDENCE READY Scoring Methodology v1.2.0 (ENGINE v1)
2026-09-05 AGIBOT A2 Ultra Batch C/D — Tiangong Ultra EVIDENCE READY Scoring Methodology v1.2.0 (ENGINE v1)
2026-09-05 Persona Humanoid Batch C/D — Tiangong Ultra EVIDENCE READY Scoring Methodology v1.2.0 (ENGINE v1)
All evidence updates

Read the methodology

Evidence tiers, operating mode, source independence, durability, and why a robot with two publishable claims can still report Very Low confidence.

How scoring works

Found an error? Tell us.

Anyone may submit a correction or new evidence, manufacturers included. Submissions never change a score directly, and scores cannot be bought.

Corrections policy