Humanoid Robot Deployments in Manufacturing Plants

Manufacturers spent most of the last decade treating humanoid robots as a research curiosity, something for conference demos and short video clips online. That changed in 2026, when a handful of programs crossed from staged footage into logged production hours on real factory floors, with named customers, contracted pricing, and measurable output. The shift is narrower than the coverage suggests, and the gap between what's actually running and what's still aspirational is what a procurement team needs to understand before committing capital.

Manufacturer attention to humanoid robots in 2026

Three things came together at once. Sim-to-real training pipelines got good enough that a robot could learn a task in a simulated environment and carry that skill into a physical plant without months of manual reprogramming. Vision-language-action models started letting a single robot generalize across tasks instead of needing a bespoke program for every motion. And actuator costs, long the line item that made legged robots impractical, kept falling. None of these three things alone would have moved the needle. Together, they turned a lab project into something a plant manager could put a number on.

A labor problem that isn't going away on its own drives all of it. An industry trade group and Deloitte have estimated that the industry needs to fill as many as 3.8 million jobs between 2024 and 2033, with as many as 1.9 million of those seats potentially going unfilled. A separate projection puts the global manufacturing labor shortage at 8 million workers by 2030. Those aren't small gaps, and they're the reason procurement teams that would normally wait five more years for a technology to mature are running pilots now.

Scale that against what's already out there. An industry trade body counted roughly 4,281,585 industrial robots operating worldwide in 2023, up 10% from the year before. Humanoids are a rounding error against that number today. But the capital flowing into the space isn't betting on the current count, it's betting on the growth curve. Market size estimates for humanoid robotics vary enormously depending on which research firm is doing the counting, MarketsandMarkets, The Business Research Company, and another research firm all publish different topline figures, with that other research firm sizing the manufacturing sub-segment alone at $7.43 billion. Treat the specific numbers as soft. Treat the direction as the actual signal: money is moving toward this category fast, even if nobody agrees on how much.

The definition of "deployed" for this analysis

The word "deployed" gets used loosely in this industry, and that looseness does real damage to anyone trying to plan around it. A robot logging hours on a factory floor could mean a research team collecting motion data for training purposes. It could mean a paid pilot with a defined end date and performance targets. It could mean a signed commercial contract with verified production hours and a price per hour of labor. Or it could mean a company running its own robot in its own facility with no external customer validating any of it.

Those are four different categories of evidence, and they carry four different weights:

Internal-only deployment means the company that builds the robot is also the only one using it, with no outside commercial terms attached. Paid pilot means a contracted, time-limited program with agreed performance metrics. Commercial contract means verified production hours, a named customer, and a disclosed price. Scale deployment means the same robot running across multiple facilities and multiple customers.

Why bother drawing these lines so carefully? Because Tesla's Optimus program and Figure's work inside BMW's Spartanburg plant get talked about in the same breath constantly, and they don't belong in the same sentence, evidentially speaking. Figure 02 logged more than 1,250 operating hours at BMW Spartanburg under a commercial arrangement with BMW, with output tied to real vehicles coming off the line. That is a fundamentally different standard of proof than internal footage of a robot picking up a bin. Conflating the two doesn't just muddy the narrative, it produces bad planning assumptions for anyone trying to build a business case around this technology.

The deployments that have crossed into verified commercial production

A small number of programs clear the bar of commercial contract or better. Here's what's actually running, with the specifics that matter for anyone trying to model unit economics.

Figure AI (Figure 02 / Figure 03), BMW Group Plant Spartanburg, USA. Figure 02 completed an eleven-month pilot at Spartanburg, logging more than 1,250 operating hours across regular production shifts, doing material handling and parts transfer. Over that run, the robots assisted in the production of more than 30,000 BMW X3s and moved more than 90,000 parts, with placement accuracy above 99%. Figure has since retired the Figure 02 units and moved Figure 03 into a logistics sequencing task at the same plant, a transition that happened in June 2026. The commercial contract now covers an initial fleet of roughly 40 Figure 03 units spread across body-shop and assembly-line workstations, priced at approximately $25 an hour. Note that the 84-second cycle time target was hit by Figure 02 during the pilot, not by Figure 03 in its current role, a distinction worth keeping straight since the two robots are doing different jobs. The Spartanburg scope keeps expanding through 2026 and into 2027, and separate pilot work is running at BMW's German plants, including the AEON deployment already underway at Leipzig. The $25-an-hour figure paired with 99%-plus placement accuracy is, as far as public data goes, the clearest unit-economics benchmark anywhere in the humanoid robotics industry right now.

Hexagon Robotics (AEON), BMW Group Plant Leipzig, Germany. This is BMW's first humanoid deployment in Europe, and it started as an initial test at Leipzig in December 2025, with a further test phase planned for 2026. AEON units, built by Zurich-based Hexagon Robotics, are working toward full production status later in 2026, handling high-voltage battery assembly and component manufacturing. AEON runs on a mobile wheelbase with a self-swapping battery system, so it can keep working across shift changes without downtime for charging. It carries 22 sensors and four layers of what Hexagon calls physical AI, including imitation learning that reportedly needs only a small number of human demonstrations to train a new autonomous task. BMW has set up a Center of Competence for Physical AI in Production specifically to pool what it learns across these programs and apply it globally, and the automaker is extending the Leipzig deployment starting summer 2026.

Apptronik (Apollo), Mercedes-Benz Berlin-Marienfelde and Kecskemét, Hungary. The partnership was announced in early 2024, marking Apptronik's first publicly disclosed commercial deployment and Mercedes-Benz's first humanoid robot application anywhere. Apollo runs intra-logistics work: delivering assembly kits, moving toted and kitted parts between workstations, and doing initial component inspections. It is not doing direct assembly. The robot itself stands 5'8", weighs 160 pounds, lifts up to 55 pounds, and runs on 4-hour hot-swappable battery packs, with a collaborative force-control system rated safe for humans working nearby. Mercedes published a case study covering six months of the Berlin pilot, reporting a 14% throughput improvement on the logistics tasks Apollo was assigned. The program is integrated into MO360, Mercedes' digital production ecosystem that uses digital twins across select plants, and Mercedes is treating the Berlin work as a blueprint it could extend across its factories worldwide. Mercedes-Benz holds an equity stake in Apptronik directly. On the manufacturing side, Apptronik has partnered with Jabil to build Apollo at scale, and a Series B extension brought the company's total funding to $935 million, with Mercedes-Benz among the investors.

Agility Robotics (Digit), GXO, Schaeffler, Amazon, Toyota Motor Manufacturing Canada, Mercado Libre. Digit's commercial deployment at a GXO facility near Atlanta in mid-2024 was the industry's first formal commercial humanoid deployment, structured as a Robots-as-a-Service arrangement rather than a straight sale. By May 2026, Digit v4 had logged more than 65,000 hours of commercial operation across nine committed customer facilities. At GXO alone, Digit moved more than 100,000 totes at roughly 98% operational accuracy. At Schaeffler, deployed across several facilities, the robot moved about 25,000 totes at a similar 98% accuracy rate. Named customers now include Schaeffler, GXO, Toyota Motor Manufacturing Canada, Amazon, and Mercado Libre, a customer roster broader than any other humanoid program currently running. Digit v4 units have passed OSHA-recognized NRTL field evaluations, a regulatory hurdle that matters more than it sounds like for anything sharing floor space with human workers. As of May 2026, Agility had more than $300 million in multi-year customer orders booked for its next-generation Digit 5, and its RoboFab plant in Salem, Oregon is designed to build up to 10,000 Digit units a year at full capacity. Agility is majority-owned by Amazon, a detail that matters given how much of Digit's early customer base overlaps with Amazon's logistics network.

Tesla Optimus: the most-watched program and the hardest to read

Optimus has the largest single-site fleet of any humanoid program by a wide margin, with an estimated 1,000 to 1,200 units deployed across Fremont and Giga Texas as of mid-2026. It's also the hardest program to evaluate against any of the standards laid out above, because none of the commercial evidence exists yet. There are no external sales, no published uptime figures, and the company's chief executive himself described the Fremont and Texas units during the Q4 2025 earnings call as being used "primarily for learning and data collection rather than performing productive tasks.""

The Q1 2026 earnings call didn't resolve much. Musk confirmed that Optimus production at Fremont would start in late July or August, but warned the initial output would be "quite slow" and called the eventual production rate "literally impossible to predict," citing the 10,000 unique parts that go into the robot. Tesla has made real physical progress on the manufacturing side: the original Model S and Model X assembly line at Fremont was reported to have been fully dismantled in a matter of weeks, and the company's Q2 shareholder update confirmed those lines were decommissioned with first-generation Optimus production lines going in, production "anticipated later this year," a genuine industrial commitment though not yet a commercial one." That's a genuine industrial commitment. It's just not yet a commercial one, and the distance between "largest fleet" and "verified commercial deployment" is the whole story with Optimus right now.

Where humanoid robots fit today

Looking across every verified deployment reveals a pattern immediately: material handling, tote movement, bin picking, kit delivery, parts transfer. These are the tasks with wide positional tolerances and a short gap between simulated training and real-world execution. A robot doesn't need to place a part within a fraction of a millimeter to move a tote from one rack to another.

None of these programs are doing high-speed welding, stamping, or high-precision electronics assembly, the tasks that actually define most automotive and electronics production lines. That's not a gap that closes soon. Traditional fixed-arm industrial robots already hit repeatability of 0.02 to 0.2 millimeters, running at cycle times several times faster than any current humanoid platform, backed by decades of mean-time-to-failure data that no humanoid program can yet match. No humanoid platform today runs at automotive-line production rates. That's a plain finding based on what's been disclosed, not a hedge.

The design logic explains why the fit lands where it does. A bipedal robot solves a real, specific constraint: it can navigate the same aisles, ramps, and clearance widths built for human workers, terrain where a wheeled autonomous mobile robot often can't go. That buys flexibility a fixed arm never had, since the same humanoid unit can get reassigned to a different job in a different part of the plant without any mechanical reconfiguration. That flexibility comes at a cost, though. A fixed arm still beats a humanoid on raw speed and precision, and for the tasks where those two things matter most, that trade-off hasn't shifted yet. The honest read for 2026 is a technology finding its lane in logistics and material flow, while the harder, faster, more precise work stays with the automation that's already been doing it for decades.