Thanks for playing! Yup, I had to compute every vertex, triangle, and uv. I'm probably going to make a video later to go in depth on the process.
With the gem collection, I cheated a bit and was relying on a quirk of how the rooms are generated. Every point in the room is visible from the center, so it was easiest to send robots back to the center after each task instead of try to do some actual pathfinding. With more time, a smarter collection strategy like you say would certainly be an improvement. But I wasn't really aiming for it to be faster than collecting it yourself, more just a way to divide your attention. When you have multiple robots, you can send one robot one way with a collector and then personally follow the other robot without having to run to the opposite side of the cavern maze.
The map is likewise very simple. Just a top-down camera view. There was no simple way to get the elevation of all the rooms, since it is determined by random slopes in the tunnels, and the rooms aren't assigned to any particular "level". The geometry isn't even stored in separated meshes. I could maybe lower the map camera to just past the last raycast hit before hitting the player, and then reduce the culling distance by casting another ray upward. I'll have to play around with that and see if that works or if it just cuts off too much of the map.