One thousand Apollo humanoids. GXO and Apptronik put that number into a letter of intent on 12 November, turning a warehouse pilot into a four-year robotics-as-a-service framework with expanded trials from 2027.

The letter is not a purchase order. Like Agility's $300M book of Digit 5 orders, it is milestone-conditioned. It also lands in a year when global humanoid shipments stayed below 100,000 units and roughly 10% of 2025 production went into real operations, according to Interact Analysis.

Our reading is blunter than the press release. The number that matters is not 1,000. It is paid robot-hours per human intervention, and almost nobody selling a humanoid publishes it.

65,000 paid robot-hours

Agility Digit 4 fleet hours

Cumulative operating hours reported across GXO, Schaeffler and Toyota sites. The figure is company-claimed, not audited. · Agility Robotics, 2026

What the spec sheet measures, and what it hides

Humanoid spec sheets have converged fast. Height sits between 1.7 and 1.9 meters, weight between 60 and 90 kilograms, payload between 16 and 25 kilograms. Boston Dynamics publishes the most complete sheet for any Western platform: Atlas carries 56 degrees of freedom, a 2.3-meter reach, 50 kilograms of instant lift and 30 kilograms sustained, a four-hour battery and an IP67 rating.

Agility's Digit 5, unveiled on 15 September, lists 22.7 kilograms of payload, a 90-minute battery and a nine-minute charge cycle. Apptronik's Apollo 2 publishes a 25-kilogram payload and roughly four hours per pack with hot-swap, though the company's own page labels weight and payload ambiguously and the numbers should be requested for the exact configuration under offer.

PlatformPayloadRuntimePublic evidence, Nov 2026
Atlas 50 kg instant, 30 kg sustained 4 h, self-swap 2026 run committed to Hyundai and DeepMind
Digit 5 22.7 kg 90 min, 9 min charge 65,000+ hours on Digit 4; $300M+ conditional orders
Apollo 2 25 kg ~4 h, hot-swap GXO LOI for up to 1,000 units from 2027
Figure 03 Not published Not published 350+ delivered; BMW pilot on Figure 02
Unitree G1 ~2 kg per arm ~2 h 5,500+ units shipped in 2025; 9.01% industrial revenue

Vendor specifications and company-claimed deployment figures, November 2026

The numbers that separate a pilot from a product do not appear in that table. Productive availability, human interventions per shift, first-pass quality and cost per accepted task are the four figures a procurement team needs. The Physical AI Journal's evidence ladder puts every named Western manufacturing programme at "piloted" or below, and none at sustained operation or independent verification.

That is the editorial problem with the category. Vendors compete on peak payload and joint count, while the buying decision turns on a denominator nobody publishes.

There is a second distortion. Joint count is a rough proxy for dexterity, not a measure of it. A 44-degree-of-freedom platform running a weak manipulation policy is a puppet with many strings and no puppeteer, as Robo2u puts it. Every actuated joint adds cost, wiring, weight and a failure mode. A procurement team comparing DOF as if it were horsepower is optimizing the wrong variable.

Actuation is the real bill of materials

McKinsey attributes 40–60% of humanoid cost to actuation, an analyst estimate relayed in the same evidence review. That single figure explains most of the trade-offs on the sheet. A joint that lifts 50 kilograms needs more torque, more copper, more cooling and a heavier structure around it. The payload number rises, and runtime, safety margin and price move the other way.

Compliance matters more than raw torque for work near people. Low-inertia, backdrivable joints yield when bumped and make force limiting credible. Stiff, high-torque joints win on repeatability and speed. Atlas and Digit sit on opposite ends of that spectrum, and the correct choice depends on whether the task is parts sequencing in a cell or tote handling on a warehouse floor.

A buyer case reported by Robo2u illustrates the trap. A logistics team chose a 20-kilogram platform over a 12-kilogram one, on the assumption that more payload meant more capability. On the floor the heavier robot cleared about 40 minutes of real work per charge, needed two people watching it and stopped on any tote outside its training set. The lighter unit ran longer and handed edge cases to a remote operator, and it moved more totes per shift.

Ask for sustained payload through a full motion at speed. The momentary lift in the launch video is a design figure.

Sensing and force feedback decide the task

Hands are where the money and the fragility concentrate. Figure's third-generation hand carries fingertip sensors that detect forces as small as three grams, which is the difference between handling a rigid box and handling a bag of groceries. Tactile sensing is still early across the industry in 2026, and "tactile hands" on a slide means little without knowing what the policy does with the signal.

Agility reports about 98% on-task accuracy across 100,000 totes moved at GXO. That is a company figure, and it describes one task in one facility. It is still the most concrete operating record in the Western field, which says as much about the evidence gap as about Digit.

Force sensing at the wrist and tactile arrays in the fingers let a controller modulate grip and detect slip. A hand that cannot feel will crush a soft item or drop a heavy one. The procurement question is not how many fingers the hand has. It is how quickly a new task can be taught and how the robot behaves when it is uncertain.

Compute placement is a latency decision

Real-time balance and control run on an onboard computer, usually a Jetson-class module. The learned perception and manipulation policies may run on a second, heavier module or partly in the cloud. That architecture choice is a deployment risk, not a footnote. A robot that needs a live connection to think will stall on a spotty warehouse network, and it raises data-governance questions that procurement teams rarely ask before signing.

Figure's Helix 2.5 raised zero-shot full-task success in home trials from 9% to 56% across 30 homes. The result is an internal evaluation, not a factory cost figure, and the company says so. It does show how fast the policy layer is moving, which is why a three-year RaaS term is also a bet on software updates the buyer does not control.

Write down what runs on the robot, what runs off it, and what the robot does when the network drops. Then write down who owns the operating data.

Safety is a gate, not a checkbox

Full-size humanoids sharing space with people remain an unsolved safety problem. ISO 25785-1, the humanoid-specific standard, sits at Committee Draft stage, and the integration part has not been developed. Vendors are applying industrial and collaborative-robot frameworks: ISO 10218, ISO 13849, speed and force limiting, emergency stops and monitored keep-out zones.

⚠️
No document, no deployment.
A 70-kilogram machine that falls carries real kinetic energy. Demand a written safety concept covering fall behavior, emergency stops, force and speed limits, keep-out zones and the supervision model. Mercedes-Benz's Apollo units are reported as teleoperated in test mode; autonomy is a stated objective, not a shipped feature.

The certification question has a commercial edge. A platform certified for a fenced industrial cell is not automatically ready for fenceless work beside untrained staff. Those are different risk cases, and the second one currently depends on a standard still in draft. A vendor that waves the distinction away is disqualified, whatever the demo shows.

The integration bill nobody quotes

RaaS pricing floats around $10–$30 per operating hour or a low-to-mid five-figure annual lease per unit, according to Robo2u's buyer guide. That range is the visible half of the cost. Task definition, environment assessment, safety review, network and charging setup, teleoperation plumbing and monitoring are buyer-entered line items that can exceed the hardware or service fee in the first year.

💡
Cost per productive hour = (acquisition or lease + integration + maintenance + supervision + tooling + energy + downtime) ÷ productive hours.
Availability moves that number as much as sticker price does. At a fixed $100,000 annual cost, moving availability from 50% to 90% cuts cost per productive hour from $50 to $27.78. Most vendors publish price more readily than availability.

Supervision is the line item buyers forget. In 2026 a deployed humanoid needs a remote operator for edge cases and on-site staff to reset, recharge and handle exceptions. A pilot that needs one supervisor per robot has different economics from one where an operator oversees five, and the contract rarely states which model the vendor assumes.

Battery strategy belongs in the same calculation. Most humanoids run two to four hours per charge against eight- to twelve-hour shifts, per McKinsey. Hot-swap packs or a charge-and-swap rotation are therefore hard requirements for multi-shift work, and the chargers, spares and floor space are part of the program.

Maintenance is the other unquoted line. Heavy actuators and batteries dominate lifecycle cost, and integrators estimate annual maintenance at 10–30% of capital cost for mid-range platforms under continuous pilot use, with premium dynamic platforms running higher. Field-replaceable joint modules shorten downtime; sealed proprietary actuators raise spare-part lead times. That difference shows up in year two, long after the pilot review has been signed.

Gartner expects fewer than 20 companies to have humanoids live in supply-chain or manufacturing production through 2028, out of roughly 200 proofs of concept. That forecast is an analyst view, and it frames the correct posture: a bounded paid pilot on one narrow, repeatable task, with metrics the buyer defines.

What belongs in the LOI

The GXO letter is a useful template because it is explicit about conditionality. Up to 1,000 units over four years, expanded trials from 2027, subject to agreed productivity and safety targets. Copy the structure, and tighten five clauses before signing.

First, define the task in one sentence, with environment and payload. Second, require paid robot-hours at a comparable site, with the number of robots, the task and the scheduled period attached. Third, set availability, intervention rate and first-pass quality as buyer-measured acceptance criteria, with targets fixed before the first shift. Fourth, state the autonomy level per task and log every assist from day one. Fifth, make the safety concept and the exit terms part of the contract, not an appendix.

None of that is exotic procurement practice. It is ordinary discipline applied to a category that has spent three years optimizing its demo reels.

The 1,000-unit letter is a bet that the denominator arrives before the fleet does. Buyers who write the denominator into the contract get paid for the wait. Buyers who do not get a demonstration with a payment schedule attached.

Humanoid robots in manufacturing: what the evidence ladder actually shows
Independent evidence review that separates company-claimed activity from verified outcomes, with a cost model and a named-programme ladder.
The best public accounting of what BMW, Mercedes-Benz and Foxconn have and have not proved.
Apptronik and GXO sign for up to 1,000 Apollo humanoids
The event anchor: a four-year RaaS framework, expanded trials from 2027, and the explicit KPI conditions behind the headline number.
Read the conditions, not the unit count. The letter is a framework, not an order book.
Humanoid robots in 2026: platforms, deployments and honest limits
Platform-by-platform comparison of Atlas, Digit, Apollo, Figure, Optimus and Unitree, with an evidence scorecard.
Useful for the specs; more useful for the gaps where vendors publish nothing.
Who is actually shipping humanoids in November 2026
A short verification pass across Tesla, Figure, 1X, Boston Dynamics, Agility and Unitree, sorted by hard evidence.
The caveat column is the valuable one: what each vendor still has not delivered.