Simulation Environment for Neural Network Dataset Generation

We present a simulation setup in the robot simulation software CoppeliaSim which is used for a synthetic dataset generation for training the neural network. In the simulator we can generate either color and depth images which can be tuned according to the real cameras mounted to the robot or robotic workplace. Vision sensors capture the simulated scene which contains different environment features, obstacles and objects of interest which can be labeled automatically with another filtering vision sensor. Except static environment which can be imported in case of known setup or generated based on height-field or simple objects. We can simulate randomly or with a specific pose oriented and positioned objects which may appear in the field of view of the robot. As an output the system produce RGB or depth information which is stored as a RGB or a gray-scale image or a combined RGBA image including the RGB data extended by depth data stored in the alpha channel. Second product of the system is a label describing different detectable classes for the neural network. The simulator is able to generate large datasets in a short period of time and produce a highly customized learning base for the neural network.