Navigation is one of the most important parts in designing and developing any intelligent vehicle. Staying operational while avoiding collisions and staying within safe conditions are priorities in autonomous vehicles. Obstacle avoidance is the core problem for mobile robots. Its objective is to allow mobile robots to explore an unknown environment without colliding with objects. The most classical approaches used in this direction relied on geometric models where local cost-maps are constructed. These methods involve low-level intelligence, without any perception process. Deep learning has been rapidly growing in recent years obtaining excellent results for many computer vision applications, such as image classification and object detection. One aspect for the increased popularity of deep learning is the ability to mitigate the need for handcrafted features. Inspired by the advantages of the hierarchical feature extraction of deep learning, this work investigates the development of a Convolutional Neural Network (CNN) algorithm to solve the problem of the mobile robot obstacle avoidance in an indoor environment. The algorithm takes raw images and robot orientation as input and generates control commands as network output. Control commands include go-straight-forward, turn-full-left, turn-half-left, turn-full-right, and turn-half-right. A dataset is collected which contains depth images (RGBD) and robot orientation data obtained by an Inertial Measurement Unit (IMU). In addition, the performance of the algorithm in terms of training options, hyperparameters, and output precision is evaluated and recommendations are provided accordingly. The final results show that the accuracy can be improved by including the robot orientation in the dataset, increasing the size of data, and tuning the network's hyperparameters. The CNN algorithm has shown great potential to get high path classification accuracy for obstacle avoidance for mobile robots.