Curriculum Learning-based Object Transportation using Region Partitioning

This paper presents a deep reinforcement learning (DRL)-based object transportation technique using a region-partitioning curriculum. Previous studies on object transportation using DRL algorithms have suffered a sparse reward problem where a robot cannot gain success experiences frequently due to random actions at the learning stage. To solve the sparse reward problem, we partition pose-initialization regions based on the distance between an object and goal, then a robot gradually extends the partitioned regions as training episodes increase. The robot has more success opportunities using this method, and thus, it can learn effective object transportation methods quickly. We demonstrate simulations to verify the proposed method.