Why Sim-to-Real Still Breaks
The policy works in the simulator and fails in the room. The gap is where the engineering lives.
Where this goes
Embodiment raises the stakes. A language model that is wrong produces a bad sentence; a robot that is wrong produces a bad outcome in a room with people standing in it.
The hard part of a robot is not the motor; it is the policy. Learned control has closed gaps that hand-tuned controllers never could, but the data to train it stays expensive and stubbornly physical.
Demonstration data is the bottleneck. Every staircase and doorknob was designed for human proportions, which is the real reason humanoid form factors keep attracting capital — they fit the data that already exists.
Sim-to-real is a tax you pay twice. The simulator that trains the policy never quite matches the friction, the lighting, the wear. Closing that gap is most of the engineering effort.