Semantic segmentation algorithms do not generalize to off-road datasets

Naively I would think rock is relatively predictable and static. Particularly somewhere with a stable climate. It is easily mapped using lidar, and has predictable grip levels.

I live in the UK and we have mud. It can be sticky like glue or just slightly slippy, depending on soil type. Puddles may have 15cm of sticky mud at the bottom or stable stones. An exposed chalk slope will be grippy, unless it has just rained.

The ideal route across the terrain changes based on lots of different factors. What is the weather like? What is the soil and geology? Is the ground free draining? Is their animal activity that has undermined the slope? Is that a stinging nettle understory or protected wild flowers?

My hunch is that most of these questions could be better answered using traditional data and methods. Use weather data, soil and geology maps, flood modelling, vertical lidar etc. Build a strategy to predict how the surface will behave under a wheel.