spatial_softmax-pytorch End-to-End Training of Deep Visuomotor Policies Deep Spatial Autoencoders for Visuomotor Learning original implementation MNIST classification using only 4 [x,y] mnist acc >> 90% feature0 feature1 feature2 feature3