Convolutional Encoder-Decoder Networks for Robust Image-to-Motion Prediction

A deep encoder-decoder network was previously proposed for learning a mapping from raw images to dynamic movement primitives in order to enable a robot to draw sketches of numeric digits when shown images of same. In this paper, the network architecture, which was previously constructed entirely with fully-connected linear layers, is modified to include convolutional layers in order to improve the image encoder component and make the network more robust to noise. The convolutional layers are pre-trained as part of an MNIST digit classifier and adapted for use in the encoder-decoder network, before the network is trained using a dataset composed of digit images and corresponding writing trajectories. This architecture was tested on several challenging noisy digit datasets and the use of convolutional layers is shown to provide a robust improvement in results.