Skip to the content
Robotics intelligence
Humanoid robotics·Capability assessment·United States, Germany, Japan

Humanoid pilots have left the demonstration cell. The throughput data has not followed.

Several platforms now work inside customer facilities rather than on a stage. What is documented is the integration, not the economics of a paid shift.

GAIA Robotics DeskPublished November 9, 2026Updated November 10, 2026v027 min readPartially VerifiedConfidence · Moderate
A tall humanoid robot standing beside idle assembly equipment in a dim industrial cell, lit by a single overhead lamp.
Generated illustrationA humanoid platform in an automotive assembly cell at blue hour.· Illustration · GAIA Robotics Desk

Executive summary

Humanoid platforms are now installed in customer facilities rather than demonstrated on stages, and the integration itself is the achievement the record supports. No party has published the throughput, intervention rate or unit economics that would establish whether the deployment pays.

The development

Over the past two years the announcements have changed character. Earlier material showed a platform walking, standing and grasping on a stage or in the developer's own laboratory. The current material shows a named customer, a named facility and a defined class of task, usually moving totes or parts between stations.

That shift matters because a customer facility imposes constraints a laboratory does not: floor space shared with people, safety zones, shift patterns, power and network provisioning, and maintenance that has to happen without stopping the line. A pilot that survives those constraints has established something real. What it has not established is the rate at which it produces work, or the cost of the supervision it requires.

Technology and capability

The platforms combine bipedal or wheeled mobility with a torso, arms and a perception stack, and are increasingly driven by learned policies rather than hand-coded motion. The dependencies that decide performance are unglamorous: battery endurance measured in hours rather than shifts, actuator duty cycles under continuous load, thermal management in a plant that is warmer than a laboratory, and the teleoperation or human intervention that is still required when a grasp fails.

The integration layer is where a pilot succeeds or fails. A unit needs a defined task, a safe approach path, a way to hand off a completed task and a way to be recovered without stopping production. Vendors describe this as orchestration; in practice it is site engineering, and it does not transfer automatically between facilities.

What has been demonstrated

Platforms have been filmed and photographed moving through customer facilities, carrying totes and moving parts, with staff present and supervision in place. The integration work is documented by the parties involved: space allocated, safety zones established, a task class agreed and a deployment period named. That is a genuine step beyond a stage demonstration and should be read as one.

What is claimed, not demonstrated

Vendors describe general-purpose capability, rapid redeployment between tasks, and a path to volume manufacturing at a unit cost comparable to a vehicle. None of the three is established by a pilot. Unit cost depends on production volume that does not yet exist; redeployment depends on site engineering that has not been shown to be repeatable; generality depends on task coverage that is not published. These are plans, and they are labelled as plans here.

Capability, as evidenced

  • Bipedal mobility inside a live production aisle

    Reported deployment

    Movement through a working facility is stated by both parties. Independent measurement of reliability in that environment is not published.

  • Dexterous manipulation of production parts

    Reported testing

    Tasks are named. Cycle time, intervention rate and defect rate are not disclosed, so the manipulation is evidenced as attempted rather than as achieved.

  • Redeployment between unrelated tasks without re-engineering

    Concept

    Asserted as the point of a general-purpose platform. No public record shows a unit moved between two unrelated customer tasks.

The organizations involved

  • Figure AI

    Platform developer

    Named by its own announcements and by the customer statement that accompanied the pilot.

  • BMW Group

    Customer facility

    The plant is named in the announcement. The scope of the work, the shift pattern and the measured output are not.

  • Agility Robotics

    Platform developer

    Named by its own material and by the logistics operator hosting the deployment.

  • GXO Logistics

    Customer operator

    The operator confirmed a deployment and the class of task. It published no rate data.

Why it matters · Lawton reading

The humanoid question has been argued as a philosophical one about the shape of a robot. It is an operations question. If a bipedal platform can be dropped into an existing facility without rebuilding the facility, it competes with retrofitting fixed automation, which is expensive and slow. If it cannot, it competes with a purpose-built machine that is cheaper and faster at one task.

The pilots are the first public evidence bearing on that fork, and they are worth reading closely for what they measure, not for how they look.

Economic consequences

Three consequences follow if the pilots hold. First, the labour comparison shifts from headcount to coverage: a platform that works two shifts and requires one supervisor changes the cost of a site, not the cost of a role. Second, the integration labour becomes a market of its own, which favours systems integrators and site engineers over platform vendors. Third, actuator, gearbox and battery supply become the binding constraint on scale, because a humanoid consumes far more of each than a fixed arm.

If the pilots do not hold, the capital already committed to general-purpose platforms does not disappear; it moves into narrower machines that do one task well, which is the pattern of the previous automation cycle.

Limitations and uncertainties

  • No party has published throughput, intervention rate or cost per operating hour for a customer deployment.
  • The deployment periods named in the announcements are short relative to the maintenance intervals of the hardware.
  • The desk has not independently observed any facility and relies on the parties' own material.

What to watch next

  1. 01A customer publishing its own measured output rather than confirming a pilot.
  2. 02A second deployment of the same platform at a different site, which would test repeatability of integration.
  3. 03Publication of maintenance intervals or actuator replacement rates under continuous duty.
  4. 04Disclosure of unit production volume, which is what the unit-cost claim depends on.

Sources · 3

  1. Deployment announcement and facility statement

    BMW Group · 2025

    Names the facility and the task class; publishes no output data.

    Original
  2. Platform material and pilot reference

    Figure AI · 2025

    The developer's own account of the platform and the pilot.

    Original
  3. Deployment confirmation

    GXO Logistics · 2025

    Confirms a hosted deployment and the class of task; no rate data.

    Original

Related intelligence