[Wingman4l7]

It's the only way this kind of robot will ever be successful. It's a bit like the driverless car approach -- get the hardware out into the real world with minimum viable performance, then desperately snaffle up as much real-world training data as you can to feed into your model, and hopefully your model will improve enough before your VC funding runs out / your product fails on the market / your product gets regulated out of existence / etc.

Simulation isn't sufficient for ML in robotics -- and they simply don't have enough training data.