And if you had any inclination as to how robotics (or indeed engineering!) works I’m sure you would know the value of testing dangerous and extremely immature equipment with people who consented to be your test subjects and in controlled environments. “Deep learning” is no excuse.

Certainly in the long run we would want to test in the real world. The argument is whether or not we are there yet. Other commenters have made a compelling argument, again based on the statistics, that we are not there yet. Do you disagree, and if so, based on what facts?

I never made an excuse nor said we should be testing these on live streets.

What I did say is that with deep learning it simply is not possible to "test" in a simulated environment to any level of certainty.

You can use simulated environments with test subjects to help develop such as system (and should do so). But you will never be able to adequately test such a system in this environment.

Why? The state space of traditional robotics and engineering problems is far more constrained. So much so that you can't even compare the fields in my opinion.

But you can still gather learning data from sensor-covered cars "in the wild" and then test it with contrived tests.