You should take a look at the research [0] people are doing using the little dog [1]. They are creating some very interesting planning algorithms for unstructured terrain like that.

0 - http://cs.stanford.edu/groups/littledog/ 1 - http://www.bostondynamics.com/robot_littledog.html

I have no idea what I'm talking about... and much less than you because I've never used drones.

When I saw the big dog prototype not falling when pushed, the assumption I made is that the eventual objective would be to just try to keep going and self-correct when unevenness is encountered.

Never thought about fully automated walkers -- just manually guided, but able self correct on rough terrain.

My education by Youtube says they're supposed to follow a person. So it might still be quite useful even blind as long as it goes roughly the route its leader went.

As far as I know the autopilot in a drone isn't so much more sophisticated than autopilots of manned aircraft. They are usually controlled by human pilots, albeit assisted with tons of artificial intelligence.

I didn't mean to imply any more than what you've described, but the key point was that you could avoid practically any obstacle by gaining altitude. Drones are often controlled by human pilots while in friendly airspace and during take-offs and landings, but who'd want to fly a series of way-points when the plane can do that itself ... Instead you switch that fancy console into game mode (just kidding).

How exactly are human pilots usually avoiding crashing into terrain/objects other than by gaining altitude?

The original point was that obstacle avoidance was not so easy for a walking vehicle ... human pilots usually try to avoid obstacles the same way (with exceptions being cases for the NTSB).