Part of path planning will involve vehicle dynamics, breaking, acceleration and steering response, and the envelope of the vehicle. All of those will heavily impact how a car should drive.

But that's something the AI can learn fairly easily, isn't it? The difficulty in this case wasn't that the AI had issues figuring out the handling of the SUV, it's that it had issues detecting a pedestrian and a dangerous situation. You can still run into these issues on a bicycle, with a much lesser chance of killing people.

This is totally true but the issue was more with the methodology of the detection rather than the detection itself. Regardless of the type of vehicle, the software wasn't good enough for real-world testing.

cycle dynamics is an unsolved control problem

That's not entirely true, I can drive a bicycle just fine.

I can drive a car just fine, too. Are we still talking about autonomous vehicles?