[allthenews]

Well, probably by training the net under such conditions. Actually I see no reason that, with a sufficiently large and accurate dataset, a neural net can't be trained to pick out a "safe[enough]" emergency landing spot. And, a well trained net should be able to fly in any condition, including limited avionics, better than a human.

You might ask where this data will be coming from. I suspect 3D physics simulations are accurate enough that they could be used to generate data. Particularly if optimized for the physics behind helicopter handling.