Apple's robot trained in 1,000 fake homes. Microagi bets the other way

Here's the line. "The models will be ready for everyone at the same time. The data won't." Munich-based Microagi puts it on its own website. It's the whole bet of a startup that raised a $55 million seed round this year. Apple just published a paper that goes straight at that sentence.

What happened

• Apple HALO: Eight Apple researchers, Vladlen Koltun among them, posted "Perceptive Humanoid Motion from Simulation" to arXiv on November 4. HALO stands for "Humanoid Autonomy, Learned in simulation, running Onboard." The training data is entirely synthetic: 1,000 procedurally generated homes, with motion from a generative model trained on human movement. No teleoperation, no capture suit.
• The result: A Unitree G1 humanoid that never saw a real home succeeded in 91% of 360 trials across six real, unscanned environments, according to the paper. It sits on chairs, walks to named objects, sidesteps through half-open doors, crawls under tables and pushes boxes. It takes spoken requests. Planning and the motion model run on a Mac mini M4 Pro in a rear enclosure, offline by default, with low-level control on the robot's own onboard computer. The code is open on GitHub.
• Microagi: Builds no robots. It records how people do physical work in factories with its own capture hardware and trains models for robots from any manufacturer. It says it passed a $100M+ annual run rate six months after launching its data platform, Shift. That figure hasn't been independently verified.
• Neura Robotics: On November 5, announced a partnership with E.ON and enviaM to bring robots into power-grid inspection. Training happens at TUM RoboGym at Munich Airport, from three sources: demonstrations by skilled workers, real robot data, and simulation on top.

Three bets on one question: where does the skill come from?

Here's what I find interesting. These aren't three robotics stories. They're a fight over who holds the owned asset in physical AI. Apple says compute. The authors spell it out: an hour of teleoperated data costs at least an hour of human time, while their corpus scales with compute. Microagi says the hands of people on the line. Neura says both, plus a sense of touch.

My rule of thumb from software: your code is no longer a moat, your distribution is. Translated to robotics: a dataset you can rebuild in simulation isn't a moat either. Apple shows that walking, sitting, dodging and navigating belong in that bucket. And it gave away the blueprint. So if your valuation rests on "we have the motion data," you just picked up a positioning problem.

The catch is in Apple's own paper

Read the limitations, not the headline. One: every new task still needs a human to write the prompt and author the key poses. Two: the robot sees but doesn't feel. Once an object leaves the frame, it no longer knows it's there, and contact-heavy tasks degrade. Three: on chained tasks, real-world success was 70%, and just 50% for three steps.

That is exactly where the two German companies are planting their flags. Neura says its Neuron AI model processes touch and sound alongside 3D vision. Microagi sells the edge cases from real shifts. To me that's the actual lesson: simulation makes the generic cheap. Value stays with whatever needs contact, accountability and process knowledge. Same as in marketing.

What I'd do now

• Capture your process knowledge. If you run production, logistics or facilities, write down and film the five moves only two veterans can do. AI doesn't fix undocumented workflows. It scales clean ones.
• Settle data rights before the pilot. Who records, and who owns the data and the trained skills? E.ON and enviaM keep control of both, according to Neura. That's the template for any contract.
• Read demos properly. 91% in a paper is strong research and a long way from a production shift. If you market with it, quote the 50% too.

Bottom line: Apple just showed that walking is getting cheap. What stays expensive is what your people know with their hands.