Where Humanoids Enter the Workplace First: Tasks and Timelines from Figure–BMW, Atlas–Hyundai and Digit–GXO

The first question a site evaluating humanoids usually asks is what other sites are actually having them do. There are plenty of demo videos, but few cases publish the customer, task, duration and metrics together. This article uses only official announcements from Figure and BMW, Boston Dynamics and Hyundai Motor Group, and Agility Robotics and GXO to set out what kinds of work humanoids took on first and at what stage, and then shows how to turn the published numbers into inputs for a cost model. It is not an investment analysis of any company.

The published cases at a glance

Sources: Figure–BMW production deployment, Figure 03 at BMW, Boston Dynamics RMAC opening, Agility–GXO agreement.

All four share one thing: the first task is picking an object and moving it to a defined place. Work that directly affects process quality, such as welding or precision assembly, does not yet appear in the published cases.

Figure and BMW: the case with the most published metrics

Figure’s November 2025 post says Figure 02 handled sheet-metal loading at BMW Plant Spartanburg for 11 months. The published values are:

• Robots arrived at the plant within 6 months, and full line deployment took 10 months.
• The robots ran a 10-hour shift, Monday to Friday.
• Cumulatively they logged more than 1,250 hours of runtime, loaded more than 90,000 parts and contributed to more than 30,000 X3 vehicles.
• Figure presented as targets an 84-second cycle time (37 seconds for loading), placement success above 99% per shift, zero interventions per shift, and placement within a 5 mm tolerance in 2 seconds (Figure post).

Because the last item is a set of targets, it should not be quoted as measured results. In June 2026, Figure 03 at BMW demonstrated a sequencing task, placing thin-walled parts into slots and pulling a large metal cart down the line, without publishing performance metrics.

Atlas and Hyundai: going through a training center

In its January 2026 announcement, Boston Dynamics said production of the product version of Atlas would begin immediately at its Boston headquarters, with the first fleets going to Hyundai Motor Group’s RMAC and Google DeepMind. The same announcement said all 2026 deployments were already committed and new customers would be added in early 2027.

According to the September 2026 RMAC opening announcement, the RMAC is a test and training center inside Hyundai Motor Group Metaplant America near Savannah, Georgia, where Atlas is learning to prepare the logistics and sequencing of parts before assembly. The same announcement laid out these plans:

1. Move in 2027 to a new building about ten times the current size.
2. Begin component assembly work in 2030.
3. Deploy 25,000 units at Hyundai and Kia plants worldwide over the next few years, and build a U.S. facility able to produce 30,000 robots a year.

What sets this case apart is that robots go through a training and validation center inside a real plant before reaching the production line. Anyone planning a rollout should budget time and cost for an intermediate stage like this.

Digit and GXO: a service contract

The June 2024 Agility–GXO announcement says the companies signed a multi-year agreement after a late-2023 proof of concept and calls it the first Robots-as-a-Service (RaaS) deployment of humanoids. At a SPANX fulfillment center, Digit takes totes from cobots and places them onto conveyors, managed alongside existing autonomous mobile robots through the Agility Arc cloud platform. The fleet size was not disclosed.

Under RaaS, the cost is a periodic fee instead of an upfront purchase, which changes the structure of any payback calculation.

What the published cases do not tell you

Just as important as what these announcements disclose is what they leave out. None of the four publishes a price or service fee, so the largest input to any payback model has to come from a quote. None reports maintenance events, part replacements or downtime, which means availability over months of operation is still unknown from public sources. Only the Figure–BMW case gives cumulative output, and even there the targets for success rate and interventions are not paired with measured values. Fleet size is disclosed only as plans (Hyundai’s 25,000 units over several years) or not at all (GXO). Before committing, add each of these gaps to your pilot agreement as something the vendor must measure and report, on the same basis you will use to compare alternatives such as conventional automation or additional staff.

Turning published numbers into cost-model inputs

The Figure–BMW case publishes cumulative totals, so a simple average is possible. With cumulative throughput $Q$ and cumulative runtime $H$, throughput per runtime hour is:

Using the 90,000 parts and 1,250 hours from the Figure post gives $q \approx 72$ parts per hour. Both figures were published as lower bounds (“more than”), so this is a rough level rather than an exact average, and because it blends idle time and shift changes it cannot be compared directly with cycle time. Dividing the 1,250 hours by a 10-hour shift gives about 125 shift-equivalents of operation.

To feed values like these into robot TCO and payback calculation, measure the following on the same basis in your own pilot:

For estimating the robot’s own availability from a spec sheet, combine this with the shift-availability calculation in how to read a humanoid spec sheet.