Target-driven Model Learning for Collision-aware Planar Object Pushing

In this paper, we introduce our end-to-end model learning approach for planar object dynamic pushing from arbitrary position to user-specified positions while avoiding collision with the surroundings. We present the learning using a target-driven deep CNN-based network and the corresponding automated dataset construction using models learned through deep reinforcement learning. We employ customized Deep QNetwork models to let the robot learn separately the greedy pushing policy with plain robot-object-target states (in simple empty pushing environment), and the local collision avoidance strategy with partial implicit local obstacle information (e.g., range data provided by laser scans). The learned models are combined as hierarchical global and local planners to automate the construction of an expert-guided database that is used to train the robot on our target-driven CNN model which maps the robotobject-target states information and local depth information to robot velocity commands. We demonstrate with simulation experiments the effectiveness of our learned models for various pushing cases and show the robustness of the supervised model to generalize to previously unseen situations.