"Our work replaces the traditional components of sensing, mapping, and planning with a single function that is represented by a neural network. This increases t

"Our work replaces the traditional components of sensing, mapping, and planning with a single function that is represented by a neural network. This increases the system’s robustness against sensor noise and reduces the processing latency." I get the feeling a bit that when they compared their depth pixel in => control out NN with the other two baseline algos , a lot of the difference is attributed to latency (one of the other algos did more temporal averaging explicitly for example). The depth input already provides a pretty good source of data for planning. The NN doesn't do anything you can't do already with the other algos, but it gets tuned automatically to a very good compromise between latency and performance, something that would be very tedious work for the designed algorithm systems and this is a strong point for building stuff like this as NNs. Article is at: http://rpg.ifi.uzh.ch/docs/Loquercio21_Science.pdf http://rpg.ifi.uzh.ch/docs/Loquercio21_Science.pdf