Using Deep Convolutional Neural Networks to Abstract Obstacle Avoidance for Indoor Environments.

In this paper, an approach to learning an obstacle avoidance program for an autonomous robot is presented. A deep learning network, which matches one that was successfully used in the past for a classification task, was replicated and used to classify ten categories in the CIFAR10 dataset. This trained network was then altered by replacing the final fully-connected feed-forward network with a new one that was initiated with random weights. Using a new database made up of images labeled with the actions taken by an operator as he remotely drove the robot, the network learned the proper action for each image. In previous work, we reported that this network operating on the actual robot successfully moved through the desired path in the training environment while avoiding obstacles. Now we have expanded this work by showing that the obstacle avoidance control program is generalized enough that it was successful when tested in three environments not seen during training.