Best Humanoid Robots: Ranked on Deployment Evidence

Most humanoid rankings score machines on the specifications vendors chose to publish and the demos they chose to film. That is a leaderboard of marketing budgets. This one applies three tests a buyer can run from outside: what a named customer has verifiably done with the robot, what the vendor admits when the admission is unflattering, and whether you can take delivery of one.

The result is uncomfortable. The machine with the strongest customer evidence is not for sale. The one with the most financial disclosure is confined to a cage. The best-resourced programme in the world describes its own deployed units as being there to learn. And the only humanoid here you can put a purchase order against carries a public, unpatched root exploit. No machine passes all three tests, so read the ranking as a map of that trade.

Companion reading: how to choose a humanoid robot, humanoid robot hardware, and the four reviews behind this piece: Figure 03, Agility Digit, Unitree G1 and Tesla Optimus.

Table of contents

1. Key takeaways
2. The ranking at a glance
3. 1. Figure 03
4. 2. Agility Digit
5. 3. Unitree G1
6. 4. Tesla Optimus
7. How to choose between them
8. Verdict
9. Frequently asked questions
10. Changelog

The ranking at a glance

1. Figure 03

Best at: proving a humanoid can hold down real production work, on the customer's word rather than the vendor's.

BMW's Spartanburg plant ran Figure 02 in its body shop inserting sheet-metal parts for the welding process, and BMW states the robot supported production of more than 30,000 X3 vehicles over ten months. BMW's own conclusion: the collaboration "demonstrated that humanoid robots can safely perform precise, repeatable work steps under real production conditions." On 25 June 2026 BMW announced Figure 03 at the same plant for complex sequencing applications in logistics. The 03 revision reads like a list written by people who have run a robot in a factory: soft components for working near people, wireless charging, tactile sensors and palm cameras in the hands, speech-to-speech audio. Underneath sits Helix, Figure's vision-language-action model, which drives hands, arms, torso and legs with one policy.

The disqualifying caveat: you cannot buy one, and the numbers most people quote come from neither company. A cost of around $25 per robot-operating-hour and a 99%+ placement accuracy figure circulate widely, appear in neither BMW's release nor Figure's, and trace to third-party analysis. Even if confirmed, a per-operating-hour cost excludes integration, cell changes, supervision, engineering time and idle hours. The verified record also runs through one customer.

Full review: Figure 03.

2. Agility Digit

Best at: disclosure. Agility is mid-SPAC, so operating data its private peers never publish sits in SEC filings, and the filings are more candid than the marketing page.

Named customers include Schaeffler (the deck photographs Digit v4 at Cheraw, South Carolina, placing 25-pound baskets of bearing components into an industrial washing machine), GXO Logistics at Flowery Branch, Toyota Motor Manufacturing Canada, Amazon and Mercado Libre. Agility discloses 98 percent accuracy on its material-handling and tote-fulfilment workflows, and 65,000 hours of operation as of May 2026. At v4's ceiling of about 16 hours of battery output per 24, that is roughly eleven robot-years across the entire customer base, and 98 percent accuracy means about one pick in fifty needs a person.

The disqualifying caveat: the robot you can deploy today lives in a cage, and the one that does not is not built. The deck lists Digit v4's safety system as "Confined to workcell". Operating alongside humans is a v5 claim footnoted to ISO 25785-1, a standard the deck itself lists as pending approval and that Agility is helping write. The marketed order book is one milestone-contingent three-year RaaS contract for 1,000 Digit v5 robots, with warrants issued to the purchaser that vest as robots deploy. No revenue figure appears in the deck, and no uptime figure has ever been published.

Full review: Agility Digit.

3. Unitree G1

Best at: existing on a purchase order. On the EDU tier it is also a genuine research platform: SDK2 in C++ with a Python interface, ROS 2 over Cyclone DDS, official robot description files, a MuJoCo model and an Isaac Lab reinforcement-learning repository, which is why it is the default machine in academic robot-learning labs. The physics are honest once you know them: about two hours of active runtime (longer marketing figures describe standby), 2 to 3 kg of arm payload, and EDU configurations that raise joint torque from about 90 N·m to 120 N·m, take degrees of freedom from 23 up to 43, and add a Jetson Orin at around 100 TOPS.

The disqualifying caveat: two of them, both knowable before you buy. The entry unit does not support secondary development, and Basic cannot become EDU later because motor mounts, wiring harnesses and compute module slots differ, so a lab that guesses wrong buys a second robot rather than an upgrade kit. Worse for anything unattended: Andreas Makris and Kevin Finisterre disclosed UniPwn on 20 September 2025, a hardcoded-key and authentication-bypass flaw in the BLE Wi-Fi configuration interface that gives anyone within wireless range root-level control. It is wormable between nearby Unitree robots, it spans the product line, and as of May 2026 independent researchers reported no patch.

Full review: Unitree G1.

4. Tesla Optimus

Best at: hardware and ambition. Gen 3 hands run roughly 50 actuators across 22 degrees of freedom on custom electromagnetic actuators using planetary roller screws, chosen for shock-load resistance during walking, torque density, precision and backdrivability. Backdrivability is the expensive property that makes force control possible and an actuator safe in contact, and it is the tell that this is serious engineering rather than styling. Gen 3 mass production began on 21 January 2026 at Fremont with Model S/X lines being converted, and the ambition (50,000 to 100,000 units in 2026, talk of 1M/year capacity) is the category's largest.

It ranks last because it fails all three tests at once. No named external customer, no way to buy one, and Tesla supplies the deflating quote itself: on the Q4 2025 earnings call on 28 January 2026, Musk described deployed units as "primarily for learning, not productive tasks". Against a 2025 target of 5,000 units, Tesla shipped hundreds.

The disqualifying caveat: the credibility damage is self-inflicted and partly dated. At the October 2024 "We, Robot" event, robots interacting with guests were human-assisted via teleoperation without clear disclosure, later confirmed by Milan Kovac as "assisted to some extent". The October 2025 demo was confirmed AI-driven, a real autonomy proof point, so citing 2024 as the current state is a year out of date. The habit worth keeping: ask of any humanoid demo whether autonomy was disclosed, over what duration, and with how many takes.

Full review: Tesla Optimus.

How to choose between them

Name which of the three tests your decision turns on.

You need a machine in your building this year and you will write software for it. The list has one entry. Buy a Unitree G1 on the EDU tier, settle the tier on the first purchase order because the second one is not a discount, and keep it on an isolated network.

You need structured material movement where a fenced cell is acceptable. Digit is the one you can sign for. Pilot it on subscription, measure your own intervention rate against the one-in-fifty implied by 98 percent accuracy, and treat every v5 number as the management expectation the deck's footnotes call it.

You are forecasting rather than buying. Watch Figure for a customer that is not BMW and for task classes beyond logistics sequencing. Watch Tesla for verifiable autonomous work at a named site over a full shift, and for sales outside Tesla.

Skip the category if you need heavy payload from a purchasable machine, unattended operation on a network you cannot isolate, a published uptime figure, or audited vendor financials. None of those exist on this list.

Verdict

Figure 03 takes first place because BMW, a customer with no reason to flatter a supplier, described what the previous generation did in production and then took the next one. That is the only evidence here that does not depend on a vendor's camera crew. Digit takes second for filing operating numbers its peers hide, and loses the top spot to a workcell it cannot leave and an order book that is one contingent contract for an unshipped product.

The G1 sits third and is still the only one most readers can own, which counts for a lot when the alternatives run on pilot contracts. Buy it with the tier boundary and the unpatched BLE exploit priced in. Optimus is last on evidence and first on ambition: real hardware at real scale, and by Tesla's own account its deployed units are there to learn rather than to do productive work. When that changes, this ranking changes.

Frequently asked questions

Which humanoid robot can I actually buy? The Unitree G1 is the only one of these four sold outright. Figure works through direct customer deployments, Tesla does not sell Optimus to the public, and Digit comes on a vendor contract, either RaaS or an ownership deal in which Arc subscription and maintenance fees keep recurring.

Which humanoid has the strongest proof it works? Figure. BMW states that Figure 02 supported production of more than 30,000 X3 vehicles over ten months in the Spartanburg body shop, and announced Figure 03 at the same plant on 25 June 2026 for logistics sequencing.

Is Tesla Optimus doing factory work? Not productively, by Tesla's own account. On the Q4 2025 earnings call on 28 January 2026, Musk described deployed units as "primarily for learning, not productive tasks".

Can any of these humanoids work next to people without a cage? Not today, per Agility CEO Peggy Johnson: humanoids "all of them right now must work inside of a safety cell until they're certified to come outside of that cell." Agility's deck lists Digit v4 as "Confined to workcell", and the v5 claim is footnoted to ISO 25785-1, which the deck lists as pending approval.

Is the Figure 03 cost per hour a real number? It is an unverified third-party estimate. The roughly $25 per robot-operating-hour figure and the 99%+ placement accuracy figure appear in neither BMW's release nor Figure's material, and such a cost excludes integration, supervision, engineering time and idle hours.

What is the security risk with the Unitree G1? UniPwn, disclosed on 20 September 2025 by Andreas Makris and Kevin Finisterre. Hardcoded keys plus an authentication bypass in the BLE Wi-Fi configuration interface give anyone within wireless range root-level control, it is wormable between nearby Unitree robots, and independent researchers reported no patch as of May 2026.