A Motion Propagation Prediction based Sim2Real Strategy Migration for Clutter Removal

Abstract

When objects are densely placed, training in the simulation with artificial samples and removing clutter are helpful to reduce the cost and risk. However, the performance of control strategy decreases in sim2real is still a challenge. This paper introduces a clutter removal method of sim2real using object motion propagation prediction. In this method, based on deep reinforcement learning, push and grasp actions are used to remove clutter. The reward of push action is calculated based on the object divergence of quadtree. The action strategy is trained in the simulation environment. Due to the position error caused by the robot pushing the object in the simulation and real environment, the object motion propagation prediction network based on graph neural network is used to predict the pushing results in the real environment and replace the real push action to training pushing strategy to improve the reward value. The pushing strategy learned in the simulation is subject to fine-tuning based on differential evolution. Compared with applying the action strategy directly to the real environment, the method in this paper has higher action efficiency and completion rate.