[marcinzm]

I suspect it's because at a certain point detailed physics matters and simulating things well enough is really hard. A robotic arm might flex just a bit, a gear may not mesh quite correctly, signals may take just a bit longer to get somewhere, a grip might slip, a plastic object might break from too much force, etc, etc.

[robotresearcher]

Sounds like a perfect domain to explore robust methods that can’t overfit to silly details.