Figure Just Built 30,000 Cars at BMW. Tesla's Optimus Built Zero. The Humanoid Robot Race Is Now a Design Problem

Source: Unsplash

The first true humanoid robot factory shift did not arrive on a Tesla stage with a fog machine. It arrived on a quiet day at BMW Spartanburg, where a robot called Figure 02 picked up a sheet metal part for the thirty thousandth time.

Two stories about humanoid robots are running in parallel in 2026, and one of them is very loud while the other is quietly winning. The summary you should walk away with is this. Figure AI's robots have already contributed to the production of more than 30,000 BMW X3 vehicles at the Spartanburg plant, working ten hour shifts five days a week, with verified placement accuracy above 99% across roughly 1,250 operating hours and 90,000 components handled. Tesla's Optimus, by contrast, has zero external customers and zero verified productive deployments as of this writing, with Elon Musk himself confirming on the Q4 2025 earnings call that current units are in "learning and data collection" mode. The interesting question is not which company has the better motors. It is which company has the better understanding of how a robot fits into the messy, social, mostly human reality of a working factory floor. That is a design problem, not an engineering one.

"Workers need to be involved in the design effort to create a product people want and love, not just a cool robot."
by McKinsey, "The next act for robotics: human machine collaboration"

I keep coming back to that line because it captures the gap between the two camps. Tesla is building a cool robot. Figure is building a coworker. Those are two different products even if they share a chassis and roughly the same number of joints.

The numbers from Spartanburg are the news the industry kept missing

Let me put the deployment data in one place because it has not been reported as a unit. Across about ten months at BMW Spartanburg, Figure 02 robots loaded more than 90,000 parts, completed about 1.2 million steps, and held the line on an 84 second cycle time target for the welding station they served. The robots removed and positioned sheet metal parts for welding, then handed off to humans further down the line. They did this on a normal industrial schedule. Five days a week. Ten hour shifts. The robots come back tomorrow morning.

That is the bar. Not "we showed it folding laundry on Twitter." Real cycle time, in a real factory, on a real product, for a real customer. BMW announced on February 27, 2026 that it is expanding the deployment to its Plant Leipzig in Germany, with a test phase starting in April 2026 and a pilot in summer 2026 covering high voltage battery assembly and component manufacturing. That is the first time Physical AI of this kind has entered European automotive production.

Now compare that to Tesla. Optimus Gen 3 production started at the Fremont factory in February 2026. Tesla is ending Model S and X production in Q2 2026 to free up the Fremont site for Optimus manufacturing. Consumer sales are now targeted for the end of 2027, a slip from earlier promises. Musk acknowledged on stage that current units are not doing useful work, that the Gen 3 design is being reworked, and that "torque density breakthroughs" rather than software are the binding constraint. None of those words mean Tesla cannot win. They mean Tesla has not even reached the starting line that Figure is already running on.

Why one team is shipping and the other is rebuilding

You can look at this two ways. Engineering people will tell you it is about hardware. Figure has its own actuator stack and a neural network policy trained directly on industrial tasks. Tesla is still chasing power to weight ratios with a more ambitious set of mobility goals. That story is true.

But the story that gets less airtime, and that I think matters more, is product framing. Figure made an explicit choice to scope its first commercial product around one customer, one cell, one cycle. BMW. Sheet metal handoff to welding. 84 seconds. That is a product. Tesla treated Optimus as a platform, in the worst sense of the word, before it had a single repeatable task. You cannot design human robot collaboration in the abstract. You design it for a specific human, doing a specific job, in a specific room.

I have shipped enough enterprise SaaS to know what happens when a team skips the specific. You get a product that demos beautifully and integrates nowhere. The robot equivalent is exactly what we are watching with Optimus. Cool video. No factory will let it on the line because nobody has scoped where on the line it goes.

The four design problems nobody on the keynote stage talks about

If you walk a real factory floor where humans and robots share a space, you start to notice that the engineering problems are downstream of the design problems. Here are the four that come up over and over, and the ones I think the next generation of humanoid products will be judged on.

• Intent communication: A human worker needs to know where the robot is about to move, what it is about to pick up, and how confident it is. Agile One, the German industrial humanoid getting attention this year, addresses this with bright colors, responsive eyes, and a chest mounted information display. That is not decoration. That is the equivalent of turn signals on a car. A two ton machine that does not telegraph intent is a two ton machine that injures someone.
• Handoff design: When the robot finishes a task and a human takes the next step, the choreography matters. Where does the robot place the part. Which side does the human approach from. How does either party signal "I am done" or "I am stuck." This is pure interaction design and it is doing more for safety than any sensor.
• Failure as a first class state: Robots will fail. The interesting question is what they do when they fail. Do they freeze. Do they back off. Do they ask for help. Do they make the human's job harder or easier. The Spartanburg deployment reportedly leaned on a "graceful degrade to a known safe state" pattern, which is a quiet but enormous design decision.
• Trust calibration over time: A worker on day one will not trust the robot the way the same worker does on day ninety. The interface needs to evolve with that relationship. Early days, the robot should be loud about what it is doing. Later, the robot can fade into the background. Most products today either stay loud forever or never speak at all.

None of these problems are exotic. We have been solving versions of them for decades in industrial UX, automotive HMI, and surgical robotics. What is new is that the failure modes are now happening at conversational speed, beside human bodies, on a moving production line.

via GIPHY

What the Tesla stage demos miss about the human side of the floor

I have walked enough factories in my career to be confident about one thing. The single biggest predictor of whether new automation succeeds is whether the people on the line believe it makes their day better. Organizations that successfully integrate humanoid robots invest in change management, upskilling, new role creation, and gradual rollout. The companies that try to drop a robot into an existing role end up with the robot getting unplugged at the end of the pilot.

This is why Figure's BMW story is so important. The robots there are not replacing humans. They are taking the most physically punishing slice of one task, the bending and lifting of sheet metal, and feeding the rest of the line. The human workers in that cell now move parts more cleanly, supervise the robot, and handle the variable parts of the workflow. BMW reported less worker strain and a meaningful reduction in repetitive injury risk on that station. That is the value, and that is the story that builds permission for the next robot to come on the floor next year.

Tesla's framing on this has been different. The pitch around Optimus has been "general purpose worker" almost from day one. That framing forces every interaction with a worker to feel adversarial. If you tell a person on a line that the robot you are bringing in can do everything they can, you are telling them their job is the goal. If you tell them the robot exists to take the worst hour of their shift, you are telling them you want them to last in this job for another twenty years. Same hardware, different product, very different reception.

This is going to keep happening in every AI native product, not just robots

The reason I am writing about humanoid robots tonight is that the same pattern is repeating in every AI native product I am seeing this year. The companies that frame their AI as a "general agent that can do anything" are the ones whose customers churn after one quarter. The companies that frame their AI as "this specific assistant for this specific job" are the ones whose customers expand their seats. I wrote about this last month at reloadux in the AI Readiness Framework piece, and again on Medium in the article on building AI native experiences, because it is the most consistent finding across the dozens of AI rollouts I have seen up close.

You can read the Figure versus Tesla story as a robotics story. I read it as the cleanest case study available right now of why scope discipline beats demo theater in AI native product design. Figure picked one cell at one factory and made it work. Tesla picked the universe and is still rebuilding the chassis. Six months from now, that gap will look bigger, not smaller, because every successful deployment compounds. Each new BMW plant Figure ships into makes the next deployment cheaper. Each Optimus generation Tesla rebuilds adds another year to the timeline.

What teams building products in 2026 should take away from this

If you are building any AI native product right now, copilot, agent, vertical SaaS, robotics, voice, the lesson from the humanoid race is the lesson the robotics industry has been trying to teach the rest of us for years. Specificity wins. Scope wins. The product that is excellent at one job will outsell the product that is mediocre at fifty.

I would add one more thing. The companies that go and sit with the humans who will work next to the AI, before they ship anything, build different products than the companies that watch their AI from a Slack thread. Figure's team reportedly spent weeks at Spartanburg before the first robot arrived, mapping every step of the welding handoff and shadowing operators. That kind of fieldwork is unfashionable and slow. It is also the reason the robot did not get unplugged.

Closing thoughts

I want to be careful not to declare Tesla dead. Tesla has won races it looked impossible to win before, and Optimus has the most ambitious roadmap in the category. Hardware is hard, the chassis rebuilds are real engineering work, and the Fremont conversion suggests Tesla is finally treating Optimus as a serious manufacturing program. None of that is wasted effort.

But the real takeaway from 2026 is that the humanoid race is not a race for the most general robot anymore. It is a race for the best designed pairing of robot and worker on a specific task, in a specific place, for a specific cycle time. Figure has shipped that pairing. Tesla has not. Until the gap between those two facts closes, the headlines will keep belonging to the team that picked the smaller, harder, slower problem and got it right.

What do you think? Drop your thoughts in the comments below. I would love to hear from anyone who has worked next to a humanoid robot, designed for one, or is currently rolling out an agentic AI product in their org. Are you seeing the same scope discipline pattern, or a different one?

Sources:
1. Figure AI, "F.02 Contributed to the Production of 30,000 Cars at BMW": https://www.figure.ai/news/production-at-bmw
2. BMW Group Press, "BMW Group bringing Physical AI to Europe, Pilot project at BMW Group Plant Leipzig": https://www.press.bmwgroup.com/global/article/detail/T0455864EN/bmw-group-to-deploy-humanoid-robots-in-production-in-germany-for-the-first-time
3. New Market Pitch, "Tesla Optimus Deployment Tracker (2026)": https://newmarketpitch.com/blogs/news/humanoid-robotics-optimus-deployment-tracker
4. Mike Kalil, "Tesla Delays Optimus Gen 3 Production for Redesign": https://mikekalil.com/blog/tesla-optimus-gen-3-production-update/
5. eWeek, "Tesla Optimus Robot Launch Timeline Targets 2027 Scale": https://www.eweek.com/robotics/tesla-optimus-robot-launch-timeline/
6. McKinsey, "The next act for robotics: human machine collaboration": https://www.mckinsey.com/industries/industrials/our-insights/the-next-act-for-robotics-human-machine-collaboration
7. Interesting Engineering, "Figure humanoid robots retire bruised after 11 months of work at BMW": https://interestingengineering.com/ai-robotics/figure-humanoid-robots-retires-bmw
8. IIoT World, "Physical AI Deployment ROI: BMW's 30,000-Car Proof": https://www.iiot-world.com/artificial-intelligence-ml/robotics/physical-ai-deployment-roi-humanoid-robots/