Humanoid robotics: beyond the viral videos
Updated: 2026-07-07
For three years, videos of humanoid robots doing backflips or opening doors dominated social networks. The winning narrative was that we were months away from seeing humanoid fleets in factories and warehouses. Reality arrived more slowly, but it arrived. During 2025 and Q1 2026, several makers finally crossed the line between demo and verifiable productive deployment.
Key takeaways
-
Figure AI, Agility Robotics, and Apptronik have moved from viral video to verifiable deployments with real industrial clients (BMW, GXO, and Mercedes-Benz).
-
Tasks that work: sheet-metal loading in pre-assembly lines, standard container handling between conveyors and shelves, parts-kit delivery alongside operators.
-
No leading maker publishes a per-unit price; the only public list price in the sector is still Unitree’s, ranging 13,500–90,000 dollars depending on model.
-
The dominant commercial model is Robots-as-a-Service (rental with maintenance included), not direct purchase, so the client doesn’t absorb the obsolescence risk.
-
The best-documented case so far, Figure 02 at BMW’s Spartanburg plant, wrapped after 11 months with more than 30,000 vehicles built and 1,250 logged operating hours.
The actors that left the lab
Three companies concentrate the best-documented industrial deployments, plus a fourth maker that competes on price rather than corporate contracts.
Figure AI had Figure 02 deployed at BMW’s Spartanburg plant (South Carolina) through an eleven-month pilot that concluded in November 2025: more than 30,000 BMW X3 units built, more than 90,000 sheet-metal parts loaded, and 1,250 operating hours across ten-hour shifts, Monday through Friday, on the task of loading sheet-metal parts onto the pre-assembly welding fixture (Figure AI[1]). With that data in hand, BMW has started testing the successor, Figure 03, at its German plant in Leipzig: initial tests in December 2025, a second round in April 2026, and a pilot phase planned for summer 2026, this time on high-voltage battery assembly (BMW Group[2]).
Agility Robotics has had Digit working since June 2024 in the warehouse GXO operates for Spanx in Atlanta, under a multi-year Robots-as-a-Service agreement that GXO itself describes as the sector’s first formal commercial deployment of a humanoid robot (GXO[3]). The task is moving containers from cobots onto conveyors; by November 2025 it had passed 100,000 totes moved in that same operation (Robotics and Automation News[4]).
Apptronik, with Mercedes-Benz as both client and investor since March 2024, has Apollo under trial at a Mercedes-Benz plant in Hungary, where it delivers parts kits to assembly workers while inspecting components at the same time (PR Newswire[5]; The Robot Report[6]). Neither Apptronik nor Mercedes-Benz has specified how many units are under trial or whether the program has expanded to other plants.
Unitree, from China, sells its G1 and H1 humanoids directly through its own store: the base G1 costs $13,500 (up to nearly $74,000 in better-equipped configurations), and the H1 starts around $90,000 (Unitree[7]; The Robot Report[8]). It’s the only one of the four makers with a public list price, which is why it serves as a market reference point even for robots that don’t compete in the same task category. As we covered in collaborative robotics on the factory floor, the pattern repeats: the high-end humanoid certified for automotive work costs an order of magnitude more than the research platform.
Which tasks actually fit
Three characteristics make a task a reasonable humanoid candidate:
-
Predictable structure: stable environment, known objects, bounded variability.
-
Duration: tasks filling full hours, not short interactions where activation cost doesn’t amortize.
-
Failure tolerance: systems where a robot error doesn’t break critical chain nor endanger nearby people.
Tasks meeting these three conditions in real plants: – Standard container handling in logistics. – Loading parts into presses and machine tools. – Visual inspection with defined routes. – Material transfer between nearby stations.
Tasks that don’t fit: variable fine assembly, ad-hoc repair, customer interaction, and anything combining non-trivial human judgment with physical execution.
Real cost and return
The real per-unit cost of Figure’s, Agility’s, and Apptronik’s industrial robots isn’t public information: none of the three makers has published a list price, and the figures circulating in trade press vary so much from one source to another that citing any single one as a hard number would be misleading. The only verifiable list price in the sector is still Unitree’s ($13,500–$90,000 depending on model), and by design that robot doesn’t compete in the same automotive-certified task category as the other three (Sacra Research[9]).
What is public is the dominant commercial model. GXO and Agility describe their agreement as the sector’s first Robots-as-a-Service contract for a humanoid robot (GXO[3]), and that same pattern, rental with maintenance included instead of direct purchase, repeats across the other announced deals. The reason is simple: no company wants to absorb the obsolescence risk of a technology that turns over generations every 12–18 months, as just happened with the jump from Figure 02 to Figure 03 at BMW itself.
ROI depends above all on the robot’s real reliability across a full shift, a figure no maker publishes yet in an auditable way. The best-documented case remains Figure 02 at BMW: 1,250 operating hours across eleven months, moving more than 90,000 parts, before the fleet was retired once the successor arrived (Figure AI[1]). It’s the sturdiest reference point that exists today for how much a humanoid actually delivers on a real production line, and it’s a better yardstick than any unverified cost estimate.
What companies that have already gone through this keep repeating is that the variable that decides the return isn’t the purchase price: it’s how many hours per month the robot runs without human intervention.
Where humans remain unbeatable
Despite enthusiasm, humans maintain clear advantage in four dimensions:
-
Fine manual dexterity, especially with soft or deformable objects (cloth, cables, unpackaged food): still out of current humanoid robotics’ reach by a considerable margin.
-
Adaptation to unexpected changes: when a belt jams in an unusual way, the robot asks for help, the human improvises.
-
Judgment on imperfect quality: deciding whether a marginal part passes or fails in an edge case is still better with an expert operator.
-
Interaction with people: a humanoid robot on-site still requires strict safety protocols and limited interaction zones.
This tension between physical autonomy and governance of the data each robot generates connects to a broader problem we covered in Industry 4.0 and data sovereignty: whoever operates the robot fleet also decides where the video and telemetry stream those robots capture, shift after shift, actually lives.
What will change soon
Three lines look reasonable in the near term, based on what the makers themselves have already announced:
-
Foundation models for motor control that several makers are training point toward behavior transfer between tasks without reprogramming, which would lower commissioning cost.
-
Fine manipulation should improve with the new-generation hands already in testing.
-
The generation-turnover pace is fast: BMW already moved from Figure 02 to Figure 03 at Spartanburg in under a year, and it’s reasonable to expect other makers to follow a similar cycle as their foundation models mature.
What probably won’t change is the problem’s nature: the general humanoid replacing the operator across all tasks remains distant because it’s not just hardware or software.
When it pays off to enter
For a company considering whether to enter humanoid robotics, three questions filter the decision:
-
Are there concrete tasks with sufficient volume (minimum one full shift per day) and repeatable structure?
-
Is physical integration possible without full line redesign or construction work?
-
Does the operations department have technical capacity to absorb the robot as a system requiring management, not passive machinery?
If the three answers are yes, it’s worth engaging with makers that have similar cases and requesting verifiable references. If any answer is no, it’s probably too early. The cost of entering too early is high and cases ending well are few.
The general lesson from the first real enterprise humanoid robotics cycle is consistent with previous tech waves: it works where well-bounded and where there’s discipline to measure outcome.
This article is also available in Spanish.
