A humanoid robot Tesla Optimus Figure 02 Unitree comparison only works if you refuse to treat the three machines as peers. Fetched on 30 September 2026, they are three different evidence classes. Unitree publishes a spec table and a shop page for the H1. Figure has published a closed BMW pilot and then a retirement notice for Figure 02. Tesla’s public robotics page still does not list height, mass, degrees of freedom, battery, or price for Optimus. A table that averages those gaps into a “winner” is a press-release mashup, not a buying document.
That is the scorecard Why 2026 Could Be the Turning Point for Humanoid Robots did not run. The April essay argued the category was moving from demo to evaluation. Six months later the evaluation is real, and it is uneven in a way that changes a procurement decision this quarter. If your constraint is a datasheet, one of these robots qualifies. If your constraint is a finished factory log, a different one qualifies, and it is already leaving the line it logged. If your constraint is a worker you can lease before year-end, none of them qualifies.
The humanoid robot Tesla Optimus Figure 02 Unitree record
Start with the document, not the demo. A spec page, a company postmortem, and a news summary of a paywalled report are not interchangeable inputs. Mixing them produces false precision: a walking speed next to a weekly output rumor next to a placement-tolerance target, as if all three numbers were measured the same way.
The filter used here is strict. A number enters the comparison only if it appears on a page opened for this piece, and it keeps the attribution that page actually supports. Unitree’s H1 parameter table is manufacturer data. Figure’s November 2025 BMW note and June 2026 Figure 03 note are company accounts of its own deployments. NVIDIA’s August 2024 unveil post is a vendor report of the Figure 02 launch configuration, not a pilot result. Tesla’s AI page is primary evidence of absence: the product description is a mission statement. Production figures for Optimus come from Embodied Wire’s 28 September 2026 tracker, which attributes them to The Information as summarized by Electrek on 25 September 2026. This host did not open The Information or that Electrek piece. Those output numbers stay labeled as reported, not as a Tesla filing.
Aggregator pages attach prices and hand counts that do not appear on tesla.com/AI or on the Figure posts opened here. Those figures are not used below.
Unitree H1 is the only public datasheet
The H1 page on unitree.com, fetched 30 September 2026, is a parameter table, not a narrative. Height is about 180 cm. Weight is about 47 kg. Battery capacity is 15 Ah, 0.864 kWh, maximum voltage 67.2 V, and the pack is described as quickly replaceable. Peak torque density is listed at 189 N·m/kg. Ultimate joint torque is about 360 N·m at the knee, about 220 N·m at the hip, about 59 N·m at the ankle, and about 75 N·m at the arm. Sensors are a 3D LiDAR plus a depth camera. Standard compute is an Intel Core i5 for platform functions and an Intel Core i7 for user development, with an optional NVIDIA Jetson Orin NX. The dexterous hand is optional.
Mobility is the line most roundups flatten. Unitree lists a moving speed of 3.3 m/s and calls it a world record, plus “potential mobility > 5 m/s.” The second figure is a potential, not a sustained run the page demonstrates. Treat 3.3 m/s as the manufacturer’s claimed demonstrated speed and the 5 m/s line as an unshown ceiling.
Degrees of freedom need the same care. The parameter table does not publish a single “total DoF” headline. It lists 5 degrees of freedom per leg (hip × 3, knee × 1, ankle × 1) and 4 per arm, marked expandable. That is 18 limb joints on the published per-limb count. Do not invent a 19th joint to make the number match an URDF you have not opened, and do not compare that limb count to a hand-heavy humanoid as if both numbers measured the same mechanism.
The H1-2 row on the same page is a different machine, and buying the wrong row is the expensive mistake. H1-2 is about 178 cm and about 70 kg, with 27 degrees of freedom, leg speed under 2 m/s, and arm joints at about 120 N·m at the shoulder and elbow. Unitree lists an arm peak load of about 21 kg and a rated load of about 7 kg for H1-2. The original H1 page leaves arm normal load blank. If the task is manipulation rather than locomotion research, the H1 row is the wrong SKU even though the brief named “Unitree H1.”
The shop listing, fetched the same day, shows $90,000 USD and titles the product “Contact us for the real price.” Duties are excluded. That is a list price with an explicit qualifier, not a transaction price and not a consumer MSRP. Unitree’s own footnote on the spec page says the global humanoid industry is in an early stage and that individual users should understand the limitations before buying. Read that footnote as part of the spec, not as boilerplate under it.
The compute line is why a lab might still want this robot when a factory would not. An optional Orin NX is a known module a team can target. That is a different problem from renting a closed policy. The practical test is the same one in Local AI Models in 2026: When Your Hardware Beats Cloud: if you cannot put your own controller on the machine, you are not buying a research platform, you are renting a demo. On-device AI in 2026 is the adjacent constraint. Edge silicon only matters if the vendor lets the policy run on it.
Figure 02 is a finished pilot, not the current robot
Figure’s own post, “F.02 Contributed to the Production of 30,000 Cars at BMW”, is the primary record for this robot. The deployment at BMW Group Plant Spartanburg ran 11 months. Robots worked a 10-hour shift, Monday to Friday. Figure reports more than 90,000 parts loaded, more than 1,250 hours of runtime, a contribution to more than 30,000 X3 vehicles, and an estimated 1.2 million steps, which it equates to more than 200 miles. The task was sheet-metal loading: pick parts from racks or bins, place them on a welding fixture. Six-axis industrial robots did the welding. This was not general assembly.
Read the KPI block as targets, not as a published scoreboard. Figure defined an 84-second total cycle, a 37-second load window, placement success above 99 percent per shift, and zero human interventions per shift. It also described the manipulation challenge as placing parts within a 5-millimeter tolerance in 2 seconds. The post does not publish a shift-by-shift series against those targets. “Minimal hardware failures” across 1,250 hours is a qualitative claim. Use the counts Figure actually totaled. Do not upgrade the targets into achieved rates.
The hardware lesson is more useful than the headline count. Figure says the forearm was the top hardware failure point at BMW. The subsystem is tight, it needs three degrees of freedom, and it runs hot. Figure 02’s forearm used a microcontroller board to distribute communications between the main computer and the wrist actuators. For Figure 03, Figure says it removed that distribution board and the dynamic cabling, so each wrist motor controller talks to the main computer directly. That is a reliability finding from hours on a line, which is the kind of evidence a demo reel cannot produce.
The same post says Figure is starting the retirement of Figure 02 and returning units from BMW to headquarters. “F.03 Arrives at BMW,” dated 30 June 2026 in the page metadata fetched with it, puts Figure 03 in Hall 52 for a sequencing job: picking unsorted parts into sequence, including pulling a cart, under a pixels-to-actions model Figure calls Helix 02. Figure 02’s body-shop pick-and-place and Figure 03’s logistics sequencing are different tasks. A sentence that says “Figure is at BMW” without the generation and the job is no longer accurate.
What Figure did not put on those two posts: height, mass, battery watt-hours, hand degrees of freedom, or a price. NVIDIA’s August 2024 unveil note adds a second onboard NVIDIA RTX GPU module, a claimed 3× inference gain versus Figure 01, six RGB cameras, and “human-scale hands.” That is launch configuration, reported by a supplier, ten months before the production totals. It is not a 2026 spec sheet, and it does not state a hand-DoF count. Do not backfill 16 degrees of freedom per hand, a 2.25 kWh pack, or a 168 cm height from a database this piece did not re-verify against a Figure primary.
The May survey The AI Hardware Race: Chips, Robots, and the Physical AI Revolution in 2026 still describes Figure 01 on a BMW floor. That generation call is stale. The robot that logged the 30,000 X3 contribution was Figure 02, and Figure’s own June 2026 post has already moved the Spartanburg story to Figure 03. The second RTX module is also an inference bill, not a benchmark trophy. AI Inference Cost in 2026 is the right ledger for that GPU: you pay for the forward pass on the robot, not for the keynote that announced it.
Helix 02, as Figure describes it on the June post, is high-frequency visual-motor control that corrects foot placement and grasp together. That is not a tool-calling agent. AI Agents Explained covers the software pattern people mean when they say “agent.” Applying that word to a pixels-to-actions policy hides the actual interface: there is no prompt you send from a laptop to re-task the BMW robot mid-shift, at least not one Figure has documented on the pages opened here.
Optimus has a ramp story and no spec sheet
Tesla’s AI and robotics page, fetched 30 September 2026, describes Optimus as a general-purpose bipedal robot for unsafe, repetitive, or boring tasks, built on vision and planning plus Tesla’s own inference hardware. It links to jobs. It does not publish height, mass, payload, degrees of freedom, battery capacity, runtime, speed, hand design, price, or a customer. Any comparison cell that fills those fields from a Tesla “press release” you cannot open on tesla.com is not a Tesla specification.
The production story is newer, and it is secondhand. Embodied Wire’s 28 September 2026 tracker says Tesla was building “several hundred” Optimus units a week in August 2026, up from “a few dozen” a week in the second quarter, citing The Information as summarized by Electrek on 25 September 2026. Managers are described as targeting more than 1,000 a week by the end of 2026 and about 20,000 a week eventually. The same summary says the robots on the line are Optimus V3 units that are not the final commercial design, and that Tesla has announced no external customers. Hands and forearms are described as more than 100 screws and small parts, still assembled manually, with a replaceable sensing glove planned for 2027. Training data is put at more than 500,000 hours, with a goal to double it by year-end.
“Several hundred” is not a precise weekly count, an internal target is not a run rate, and a glove planned for 2027 is not a shipped sensor. This host did not open The Information or the Electrek summary. If those pages disagree with the paraphrase, the paraphrase loses. Until Tesla publishes a weekly figure, the production cell is “reported ramp, unaudited.”
The unveil gap is the other half of that cell. Embodied Wire says Musk, on the January 2026 call, put a Generation 3 unveil in the first quarter, then on 22 April 2026 moved it to “probably middle of this year.” As of 28 September 2026 that reveal still had not happened, even though V3 units were already coming off the line. Output before a public spec is a factory fact. It is not a product a buyer can inspect.
The same tracker attributes about 10,000 unique parts to Musk’s April comments, plus a Fremont Model S and Model X line conversion aimed at late summer. It also says early go-to-market is leasing, and that Musk has spoken of sales beginning in late 2027, via Gizmodo. This host did not open that Gizmodo piece. A late-2027 comment is not a 2026 lease schedule, and no external customer is named in the tracker text fetched here.
Older “$20,000 to $30,000” talk does not become a 2026 price because a database still prints it. Tesla has not published a price on the page fetched for this piece. Printing one would be the same error as printing a Figure 02 price that figure.ai did not state.
The table you can defend
Empty cells are the point. A dash means the opened primary did not say, not that the value is zero.
| Field | Tesla Optimus | Figure 02 | Unitree H1 |
|---|---|---|---|
| Evidence class | Mission page plus a reported ramp | Company pilot postmortem | Manufacturer spec page |
| Height / mass | — | Not on the posts opened | About 180 cm / about 47 kg |
| Speed | — | Not published as a walking spec | 3.3 m/s listed; >5 m/s is potential |
| Battery | — | Not on the posts opened | 0.864 kWh, swappable |
| Hands | Reported manual assembly, 100+ small parts | Forearm was the top line failure; 3 DoF in that subsystem | Dexterous hand optional; arms 4 DoF |
| Compute | Tesla inference hardware, no model card | Second RTX module at unveil (NVIDIA, Aug 2024) | i5/i7 standard; Orin NX optional |
| Deployment | No external customer named | BMW body shop, then retirement | Sold as a research/industrial platform |
| Public price | None on tesla.com/AI | None on the posts opened | $90,000 list; shop says ask |
| Status on 30 Sep 2026 | V3 reported in production, not unveiled | Previous generation; Figure 03 is the live BMW robot | Spec page and shop page both live |
You cannot rank speed: only Unitree published one. You cannot rank value: two robots have no public price, and the third says the sticker is not the price. You can rank what a stranger can audit without an NDA. On that axis the order is the H1 spec page, then Figure 02’s pilot log, then Optimus. That is not a product ranking.
Hands are the constraint, and they are not the same hands
All three stories stall at the distal limb. They stall for different reasons, which is why a single “dexterity” column is useless.
On the H1, the hand is an option. The robot Unitree specifies in the base row is a fast, light biped with 4-DoF arms and a LiDAR. Ordering it to do Figure’s sheet-metal job would be a category error. The H1-2 arm ratings exist precisely because Unitree split locomotion and manipulation into two rows. Use the split.
On Figure 02, the hand-and-forearm stack did real work and then became the reliability limit. The useful result is not “human-scale hands” from the 2024 NVIDIA post. It is a named failure mode after 1,250 hours: a distribution board in a 3-DoF forearm, removed in the next generation. That is the sentence a manufacturing engineer can use. A DoF count from an unopened press release is not.
On Optimus, the hand is a production bottleneck before it is a capability claim. Embodied Wire’s summary says the hand and forearm still require manual assembly of more than 100 small parts, and that a replaceable sensing glove is planned for 2027. If that paraphrase holds, Tesla is scaling a robot whose most failure-prone subsystem is not yet on an automated process. Figure already learned that lesson on a customer line and redesigned the wrist electronics. Tesla has not published an equivalent postmortem. Volume without that postmortem is not evidence the hand is solved. It is evidence the factory is willing to absorb manual rework while it counts units.
Do not turn Figure’s 10-hour BMW shift into a battery rating, or Unitree’s 0.864 kWh pack into a runtime the spec page does not print. The shift is a schedule. The pack is swappable, which is the design answer to a small battery, not a measured hour count. Optimus has neither number in the sources opened here. That ledger is not the grid bill in AI Energy Consumption 2026.
What to do this quarter
If you need a robot you can specify and order for locomotion research, the H1 is the only name in this brief that clears the bar. Budget past $90,000. The shop page says the real price is a sales conversation, and duties are extra. Decide before you write the purchase order whether you need a hand. If you need a rated arm load, read the H1-2 row and price that machine, not the 47 kg runner. Confirm the optional compute. An Orin NX that is “optional” is not in the crate unless the quote says so.
If you need a factory reference, read Figure’s November 2025 post and stop. The counts are usable: 90,000 parts, 1,250 hours, 30,000 X3s, a forearm that failed first, a next generation that deleted the failure’s board. Do not cite Figure 02 as the robot currently on the Spartanburg line. Figure’s 30 June 2026 post says Figure 03 is the robot in Hall 52, on a sequencing task, under Helix 02. That robot is outside this brief’s product name. Comparing it as if it were Figure 02 would repeat the Figure 01 error already sitting in the May hardware survey.
If you need a worker you can lease in 2026, none of these three is a procurement answer. Optimus has no named external customer in the September reporting, the V3 reveal had not happened as of 28 September, and a late-2027 sales comment is not a delivery slot. Figure 02 is being retired. The H1 is buyable and is not a drop-in associate for an unstructured shift. Waiting for a three-way bake-off will not create the missing customer, the missing spec sheet, or the missing hand option.
The concrete next step is a three-document check, not another demo. Open unitree.com/h1, Figure’s BMW postmortem, and tesla.com/AI. If a comparison you are about to forward contains a number that is not on one of those pages or in a report you have actually read, delete the row. Then pick the constraint you really have — datasheet, pilot log, or lease — and buy, cite, or wait. Do not average them.
FAQ
Can I buy a Tesla Optimus in 2026?
Not from any page Tesla has published that this piece opened. tesla.com/AI describes the robot and links to jobs. It does not list a price, a configurator, or a customer. Embodied Wire’s 28 September 2026 summary of The Information says Tesla has announced no external customers, that early units would be leased to selected companies, and that Musk has spoken of sales beginning in late 2027. Treat 2026 as a factory-ramp year, not a purchase year.
Is Figure 02 still working at BMW?
Figure’s November 2025 post says it is starting fleet-wide retirement of Figure 02 and returning units from BMW to headquarters. The 30 June 2026 post says Figure 03 has arrived at Plant Spartanburg for logistics sequencing, after Figure 02’s 2025 assembly-line deployment. Cite Figure 02 as a completed body-shop pilot. Cite Figure 03 if you mean the robot Figure says is on the floor now.
Is the Unitree H1 the same robot as the H1-2?
No. On the spec page fetched 30 September 2026 they are separate columns. H1 is about 47 kg, with 4-DoF arms and a listed 3.3 m/s. H1-2 is about 70 kg, 27 degrees of freedom, speed under 2 m/s, and a rated arm load of about 7 kg. The brief named the H1. If the job needs hands, price the H1-2 or an H1 quote that explicitly adds the optional dexterous hand, and do not assume the base row includes it.
Which of the three is fastest?
Only Unitree published a speed, 3.3 m/s, on the page opened for this piece. Figure did not publish a walking speed on the BMW posts. Tesla did not publish one on tesla.com/AI. A three-way speed ranking is not available. Unitree’s “potential mobility > 5 m/s” is not a second measured result.
Why not pick a winner?
Because the brief’s three names are not substitutes. The H1 is a specified, orderable locomotion platform with an optional hand and a qualified $90,000 list price. Figure 02 is a retired pilot with a real hour count and no public price. Optimus is a reported production ramp with no public spec sheet and no named outside customer. The winner depends on whether you need a datasheet, a pilot log, or a lease. Collapsing that into one score hides the only decision the sources support.