Autonomous Robotic Arm Object Grasping through Consistent Depth Estimation

This thesis propose an innovative method for the estimation of the depth in a consistent means, focusing on a grasping task scenario. Infer depth informations from the environment through sensors, such as Kinect or a stereo camera, results difficult for tasks that involves robotic manipulators. Beside the usually high cost of these sensors, it is also challenging mount them on the manipulator end effector, due to their weight and their significant dimensions. Nowadays, a popular solution is monocular Depth Estimation, whose goal is to predict the depth value of each pixel, given only a single RGB image as input, allowing to infer scene geometry from 2D images. However, their quality is limited by the ill-posed nature of the problem and the lack of high quality datasets. Recent state-of-the-art methods, such as Structure-from-Motion and Multi-View Stereo, are able to generate consistent depth values but, differently from monocular depth estimation, they also require a huge number of images to do so. This mean more time spent gathering environment informations and consequently more time spent to perform the overall grasping procedure. The method proposed in this thesis, aim to reduce the amount of data needed to predict depth values, without losing consistency and reliability on the predictions. Firstly, a monocular depth estimator is used to predict raw depth values, who will be refined by a second neural network. The core idea, is the implementation of a particular Autoencoder Neural Network, whose loss function is computed by a warping procedure based on the predicted raw depth values itself. Lastly, a Salient Object Detector, is used to remove outliers from the refined depth values provided by the autoencoder. In order to test and validate the proposed method, an UR5 robotic manipulator along with a Kinect camera, has been implemented in the CoppeliaSim simulator.