Figure 3
Visualization of the behaviors of the actor robots and Observer network architecture (A) We pre-programmed four types of behaviors for the actor robot. The images shown here are produced by integrating a sequence of frames from a video captured by the top-down camera. The robot path is shown in black, the rectangular obstacle (if any) in red and the goal circles in green or red. (B) The image prediction network is composed of several layers of convolutional units and deconvolutional units. At the deconvolutional stage, we utilize multi-scale prediction to maintain high resolution at the output image. Numbers indicate the dimension of output feature map after each module.
