Thin objects are really hard, and are probably the single area we have made the most software progress since the R1, especially very thin twigs and branches (it's still not perfect, fly carefully). We completely rebuilt the stereo vision algorithms to try to handle those better, basically by incorporating semantic knowledge (recognizing what is a tree, or a power line) in addition to traditional stereo vision techniques. To my knowledge, our systems are significantly more advanced than even the state of the art in academic research papers, so I would be impressed if there are any other drones that match our performance today.
