State-of-the art robot learning approaches are data-driven and require a large amount of diverse robot data which are prohibitively expensive to acquire. In this paper, we present PickSim, an open source dynamically configurable Gazebo-based pipeline for the purpose of robot manipulation learning. Specifically, we propose a new plugin and pipeline, based on the well known open source robot software Gazebo. We showcase the potential of PickSim through the generation of 6D pose estimation dataset, potentially randomized over a number of factors (e.g., lighting, texture, shape, etc.). Rich and perfect annotations (e.g., object masks, poses, occlusions, etc.) are automatically generated and enable the learning and evaluation of computer vision models over a number of varying factors. PickSim unlocks multi-task robot manipulation learning when a robot model and physics engine are integrated into it, and thereby paves the way to the generation of large-scale robot manipulation data as required in general-purpose robot manipulation learning.