Yeah that's what I meant. Fix the nag and the bias to the right and things would be great.
Yeah, I tested this out a bit more in places where I know it would have insufficient input.
Surprisingly, the path planning seems to work rather well even with no lane markers or a vehicle ahead. Confidence is very low, and it drives a little like a drunk, but it seems to know where the road is at least and stays on it. Once it loses the road entirely (as in, trying to drive through a T-junction or something) it still gives the "TAKE OVER IMMEDIATELY" thing. However it's definitely much less likely to jump to that with the hands-on requirement disabled.
There is a road nearby that has full lane markings, then they vanish as it turns into a slightly narrower private paved road with no markings. Asphalt road, grass on sides. At the end of that road (maybe a quarter mile) the road T's in both directions. The car will drive fine up and on to the unmarked portion and continue to drive on that stretch. It does somewhat center itself in the road at this point, mostly, but will bounce around a little bit as the road curves. It seems to recognize where the road edges are and stays within them even with no markings. The IC display shows the lighter road area, but no lanes, and it jumps around a bit on the screen. When I get to the end of the road where it T's, the car tries to turn slightly towards whatever side it's already on, then gives up when it can't see any more road with "TAKE OVER IMMEDIATELY". I tried this a couple of times at maybe 20 MPH. In normal mode it tells me to take over immediately when approaching the unmarked portion.
Another example is an overpass on the interstate nearby that has no markings whatsoever. The road goes from asphalt to unmarked concrete that extends maybe 150 ft. Normally the car immediately fails with "TAKE OVER IMMEDIATELY" here consistently. In hands-on-disabled mode, it continues on mostly straight, then latches to the lane on the other side a few feet from the end of the overpass and does a kind of snap-to-grid motion to center up.
So it seems the confidence level in hands-on and hands-off modes are a bit different.