So last summer we ported the indoor experiment on to the android platform and carried out the outdoor navigation experiment.
The motivation is that the network needs to perform on-line learning with no cloud back end (to support real-time response).
Although CNN can do real-time forwarding pass with modern libraries like Caffe, but training in real-time on resource limited platform is still too much to ask.
A major advantage of our approach is that we are using competitive hebbian learning so that only a small portion of the network is updated according to the current learning input.
Detail about the experiment can be found in this paper (accepted by CVPRW, 2016).
Video demonstrations are also available on my youtube channel (yes I have an youtube channel :))
Demo1: Real-time testing
Demo2: Training and testing
Demo3: Blind testing