Humanoid Robots in the Real World: Inside the BMW and Mercedes Factory Experiments

If you still think humanoid robots are sci‑fi props, the car industry would like a word.

In South Carolina and Berlin, life‑size robots that look vaguely like people are testing the idea that “general‑purpose” machines can do real factory work: hauling sheet metal, delivering parts, and taking over the kind of awkward, repetitive jobs humans hate and often get injured doing. This isn’t a glossy concept video. BMW and Mercedes are running multi‑month pilots on live production lines, with real cars and real risk.

For you – whether you sit in operations, IT, or the C‑suite – these experiments are a preview of what your plants and warehouses could look like in just a few years. But it’s also easy to get lost in the hype. Are these robots actually productive? Are they safe? How “AI” are they, really? And do they replace traditional industrial robots, or just add another maintenance headache?

Let’s unpack what BMW and Mercedes are actually doing, how it works, and what you should be watching for next.

Why Humanoid Robots in Factories, Anyway?

Factories already have plenty of robots. So why drag in something with legs?

The pitch from companies like Figure and Apptronik is simple: a humanoid robot can, in theory, work in spaces and with tools designed for humans, instead of forcing you to redesign your lines around machines.

BMW and Mercedes are interested in humanoids mainly for jobs that are:

• Repetitive and ergonomically bad – e.g., awkward lifts, overhead work, or constant bending.
• Variable – where things change too much for a simple fixed robot arm.
• Hard to staff – shifts or tasks that humans don’t want to do for long.

BMW’s trials at its massive Spartanburg plant (the company’s only US vehicle plant) focused first on body shop tasks like loading sheet‑metal parts into fixtures, the sort of work that’s boring, tiring, and a magnet for musculoskeletal injuries. In later phases, humanoids moved into logistics and sequencing, where parts need to be picked and delivered in just the right order for final assembly. BMW has publicly described this as part of its “Physical AI” push, integrating computer vision and AI planning into the production system rather than just dropping in a robot arm and calling it a day. BMW Physical AI and Figure 03 project

Mercedes, meanwhile, is using humanoids for intralogistics: moving components and kitted parts from storage to the line, and inspecting them on the way. Their goal is to free up human workers from the most physically demanding, low‑value tasks, while proving that humanoids can actually coexist with humans on the floor without causing chaos. Apptronik–Mercedes Apollo agreement

The big bet: it’s cheaper and faster, long‑term, to drop in “human‑shaped” robots than to keep re‑engineering factories every time workflows evolve.

Case Study 1: BMW and Figure’s Factory Humanoids

BMW’s humanoid journey started publicly in early 2024, when it signed a commercial agreement with California startup Figure AI to deploy humanoid robots at its Spartanburg plant in South Carolina. The plant builds SUVs like the X3 and X5 and is one of BMW’s most important global sites. TechCrunch coverage of the BMW–Figure deal

From trial to real production

BMW and Figure have gone through several generations:

• Figure 02 body shop test: BMW reported that the Figure 02 humanoid successfully loaded sheet‑metal parts into fixtures in the Spartanburg body shop during a multi‑week test, stepping in for tasks that are awkward and tiring for humans. BMW press note on Figure 02 trial
• Extended deployment: According to BMW and later press, a Figure robot then supported production on the X3 line for roughly 10–11 months, working production shifts (10 hours a day, 5 days a week) and contributing to more than 30,000 BMW X3 vehicles by handling repetitive sheet‑metal loading. Time Magazine report on Figure at BMW
• Figure 03 logistics pilot: In 2026, BMW announced the Figure 03 robot arriving in an assembly and logistics hall (Hall 52) at Spartanburg, targeting “complex sequencing applications” – essentially, fetching and delivering series‑specific components in the right order, and adapting to dynamic line conditions. BMW Figure 03 project details

Behind the scenes, these robots are running what Figure calls “Physical AI” – a stack of vision, motion planning, and task‑level control. Earlier on, Figure even had a research partnership with OpenAI to integrate large language models, though the company later shifted focus to high‑rate robot control rather than chatty interfaces.

Where humanoids actually fit in BMW’s stack

It’s important to understand what humanoids are not doing for BMW:

• They are not running the entire line.
• They are not doing precision welding or painting (classic fixed‑robot territory).
• They are not “replacing workers wholesale.”

Instead, they are:

• Taking on single, well‑scoped workflows like a particular sheet‑metal load or a logistics sequence.
• Working within BMW’s iFACTORY strategy, which already uses AI for visual and acoustic quality checks (BMW calls this AIQX) and digital twins to plan the line. The humanoid is basically another endpoint in that AI‑driven system, not an isolated gadget. BMW on AIQX and Physical AI

From BMW’s perspective, the win is twofold: de‑risking hard human jobs today and learning how “general‑purpose” robots could keep lines flexible tomorrow.

Case Study 2: Mercedes and Apptronik’s Apollo

On the other side of the Atlantic, Mercedes‑Benz has partnered with Austin‑based startup Apptronik to pilot the Apollo humanoid robot in its manufacturing network.

In March 2024, Apptronik announced a commercial agreement with Mercedes to deploy Apollo in selected plants. The public focus: logistics use cases such as delivering parts to the line, bringing totes of kitted components to workers, and inspecting those parts en route. Apptronik press release on Apollo at Mercedes

Where Apollo works today

Based on Mercedes and industry coverage:

• Berlin Digital Factory Campus: Mercedes has been testing Apollo at its historic Berlin‑Marienfelde site, which the company has turned into a digital and AI innovation hub. Here, Apollo supports logistics tasks – moving components and assisting workers on the line – while Mercedes layers in AI tooling like its MO360 production system and AI‑based analytics. Autocar coverage of Mercedes AI robotics
• Other plants and pilots: Industry reports describe Apollo being tested in intralogistics roles such as line‑side material movement at Mercedes sites, including transporting components or modules to the production line in Hungary and Germany. Automotive Logistics report on Apollo

Apollo is designed to be tall enough (around human height), to carry meaningful loads (tens of kilograms), and to work alongside people with built‑in force control for safe physical interaction. In many cases, it doesn’t even need its legs – Mercedes can mount the torso on a base if the job is stationary.

What Mercedes is trying to learn

Mercedes has framed Apollo as part of a broader transformation of its production network, pairing:

• AI software (including work with Google DeepMind and custom large language models to assist staff).
• Humanoid hardware (Apollo) for the dirty, dull, and dangerous bits of manual work.

Their goals echo BMW’s:

• Reduce physical strain on workers.
• Increase flexibility when line layouts and product mixes change.
• See if one robot platform can be reused across plants and workflows, instead of buying a dozen niche machines.

So far, what’s been announced are pilots and early deployments, not a 10,000‑robot fleet. But the direction of travel is clear: humanoids are being treated as serious tools, not just PR props.

What’s Actually “AI” About These Robots?

If you use tools like ChatGPT, Claude, or Gemini, you’re used to AI that lives in text boxes. Humanoid robots bring AI into the physical world, but the ingredients are familiar.

A typical factory humanoid combines:

• Perception AI: Camera‑based systems that identify parts, tools, shelves, people, and obstacles. Think of it as a real‑time, embodied version of image models like Gemini or GPT‑4V.
• Motion and grasp planning: Algorithms that figure out how to move arms, hands, and legs without colliding with the environment – similar in spirit to a pathfinding engine in a game, but in 3D continuous space, under gravity.
• Task‑level control: Systems that know the steps of “pick part A from bin B, orient it like this, place it in fixture C” and can adapt when something is slightly off.
• Learning from demonstration: Many humanoids are trained partly by humans teleoperating them or guiding them through tasks, then using offline learning methods (often cousins of the reinforcement learning or imitation‑learning techniques you see in research) to generalize.

Language models do show up, but usually behind the scenes – for diagnostics, natural‑language instructions, or connecting high‑level planning. You might see robots controlled by internal copilots powered by models similar to ChatGPT, Claude, or Gemini, but you’re not going to have a production robot casually taking arbitrary prompts from the open internet.

The Benefits – and the Fine Print

From BMW and Mercedes’ early experiments, some themes are emerging that matter if you’re considering humanoids down the line.

Potential benefits

• Faster deployment: In theory, humanoids need less custom hardware integration than traditional industrial robots because they can use the same space, tools, and racks as humans.
• Ergonomics and safety: Offloading heavy, awkward, or highly repetitive tasks can cut down on injuries and long‑term health issues.
• Workforce flexibility: Humanoids can help cover labor shortages on nights or weekends and adapt to new tasks with software updates rather than full mechanical re‑engineering.
• Data exhaust: Every motion generates data you can feed into analytics and simulation tools, letting you optimize processes in ways that are hard to do when humans are the only actors.

But the fine print is real

• They are not plug‑and‑play. Each workflow still needs scoping, mapping, safety assessment, and integration into your MES/ERP systems.
• Reliability is unproven at scale. A robot working well for 10 months in one plant is promising, but not the same as thousands of units across dozens of factories.
• Safety standards are still evolving. There is no single global safety standard for humanoids yet; integrators and insurers will be cautious until best practices solidify.
• Change management is huge. You’ll need to bring operators, unions, and safety officers into the conversation early. BMW and Mercedes emphasize that robots are there to “support” their workforce – but internally, that only works if people trust the rollout.

What This Means for You (Even If You Don’t Build Cars)

You might not run an automotive plant, but BMW and Mercedes are effectively acting as high‑budget test labs for the rest of industry.

If you’re in manufacturing, logistics, or large‑scale operations, their experiments hint at what’s coming:

• Phase 1 (now): Pilots in large enterprises with deep automation experience and R&D capacity.
• Phase 2 (next 3–5 years): Broader pilots in tier‑1 suppliers, large warehouses, and logistics hubs, as vendors productize what worked in those early trials.
• Phase 3 (later this decade): Down‑market offerings where mid‑sized factories can subscribe to a few humanoids as a service, much like you lease AMRs or cobots today.

Meanwhile, the software side is maturing fast. The same AI backbone that powers ChatGPT, Claude, or Gemini-style copilots for your office staff is increasingly being adapted to control physical systems, from warehouse robots to, yes, humanoids.

You don’t need a humanoid pilot next year. But you do need a strategy for:

• Where physical AI could offload your worst human tasks.
• How your data infrastructure will support robots that constantly sense and log everything.
• How your workforce will be retrained into higher‑skill roles around supervision, troubleshooting, and optimization.

How to Get Ready for Humanoid Factory Pilots

If you want to be ready when vendors start pitching you their own “Apollo” or “Figure” for your site, here are concrete steps you can start now:

1. Map your pain‑point workflows. Identify 3–5 tasks that are:

   • Physically demanding or injury‑prone,
   • Highly repetitive but not easily solved with a fixed robot,
   • Constrained to a clear, bounded area of the plant.

2. Clean up your digital house. Humanoids will plug into your existing AI and data stack. Investing now in:

   • Good sensor coverage (cameras, scanners, IoT),
   • Clean, accessible production data,
   • Familiarity with AI tools (from analytics to copilots like ChatGPT or Gemini), will make later robot pilots far smoother.

3. Start a people‑first automation conversation. Talk with operators, health and safety, and HR about where they’d most like to see high‑risk work automated and what upskilling paths could look like. The tech is advancing quickly; your cultural readiness will be the real bottleneck.

Humanoid robots in factories aren’t magic, and they’re not mature at mass scale yet. But BMW and Mercedes are showing that they can do real work, on real lines, right now. If you start preparing your workflows, data, and people today, you’ll be in a position to adopt them on your own terms – instead of scrambling to catch up when they arrive at your dock door.