Robotics has discovered an infinite energy source: claiming your robot is “simulation-first” while quietly hiring a small army of humans to generate the data that makes the simulation useful.
In the same week Siemens bragged that simulation-first development compressed a humanoid prototype timeline from 18 to 24 months down to about seven, MIT Technology Review reminded everyone how robots actually learn: rules, then reinforcement learning in sim, then foundation-model era data hunger, then the pragmatic approach, deploy early, collect data, and improve.
The hype version of this story is “simulation replaces reality.” The real version is better: simulation is becoming an engineering workflow, and humans are still the scaffolding that makes it work.
The myth
Myth: Robots can learn everything in simulation, and then simply transfer to the real world like it’s a downloadable skill pack.
The reality
Reality: Simulation helps you iterate faster, but it cannot fully model the ugly details that break robots, contact physics, friction variability, sensor noise, misaligned fixtures, and the many ways the world refuses to match your CAD file.
So the field does what it always does: it narrows the gap with data. Sometimes that means domain randomization. Sometimes it means recording humans doing tasks with controllers, and building datasets the robot can imitate. Sometimes it means deploying systems that are “good enough” purely to harvest the error cases.
The non-obvious thing
“Simulation-first” is not a claim about autonomy. It is a claim about iteration speed.
If you can prototype mechanics, train policies, and validate behaviors without waiting for physical robots to break (or for your factory manager to stop glaring at you), you can move faster. That is valuable even when the robot still needs human exception handling on the floor.
In other words: simulation is turning robotics into something more like software engineering. But reality still gets the final merge approval.
The Droid Brief Take
I am pro-simulation. I am also anti-pretending.
If your press release says “simulation-first” and your deployment still requires a human nearby with a rescue plan, that’s not a scandal. That’s how engineering works. The scandal is when companies sell that scaffolding as “general-purpose autonomy,” and then act surprised when customers ask for intervention rates and safety validation.
What to Watch
- Evaluation, not demos. Simulation is most useful when it becomes verification, repeatable tests that correlate with real-world failures.
- Force control progress. If tactile sensing and control remain weak, sim will keep producing policies that look great until the first unpredictable contact.
- Data pipelines. Who can gather, label, and learn from real-world mistakes the fastest, without turning every deployment into a human babysitting service?
Sources
Siemens — “Siemens and Humanoid bring Physical AI to the factory floor”
MIT Technology Review — “How robots learn: A brief, contemporary history”