Humanoids Win Through Deployment Data

The moat in humanoids is shifting from money and prototypes to learning loops in live workplaces. Figure has raised $1.75B and reached an active BMW line, Tesla has unmatched road data but not indoor factory data, and the common bottleneck is still the same, getting robots into real jobs, catching every failure, and feeding those edge cases back into the model through teleoperation and retraining.

• Figure is one of the clearest proofs that deployment matters more than slideware. Its Figure 02 robots ran 10 hour shifts at BMW Spartanburg, handled 90,000 plus parts, and contributed to production of 30,000 plus vehicles. That kind of repetition creates the labeled manipulation data humanoid models need.
  2 figure 5 bmwgroup
• Tesla still has the biggest systems advantage if humanoids become a manufacturing scale problem. It can fund robots off the auto business and deploy first inside Tesla, SpaceX, and Boring facilities, but road video does not automatically teach a robot how to grasp sheet metal, carry bumpers, or move through crowded indoor workflows.
  1 sacra 4 sacra
• The rest of the field is building different shortcuts around the same data scarcity. Apptronik is paired with Google DeepMind on Gemini Robotics, Agility has focused on narrow warehouse tasks and live tote handling, and several companies are using teleoperation or synthetic data to accelerate learning before they have large fleets in the field.
  1 sacra 6 deepmind 7 sacra 8 sacra

Over the next few years, the winners are likely to be the companies that turn a few narrow factory tasks into thousands of daily repetitions, then expand outward from there. Once enough indoor manipulation data is accumulated, humanoids stop looking like custom pilots and start looking like a compounding software and fleet business.