GTHSL: A Goal-Task-Driven Hierarchical Sharing Learning Method to Learn Long-Horizon Tasks Autonomously

Rong Jiang, Xu Cheng, Hongrui Sang, Zhipeng Wang, Yanmin Zhou, Bin He

Abstract: Robot skill autonomous learning is a long-standing goal in robotics. However, for long-horizon tasks common in robotic applications, it is very challenging to achieve autonomous learning from scratch due to the long decision horizon and the high-dimensional continuous search space. In this article, we consider a multigoal pick-and-place task within obstacle scenarios, which is a typical long-horizon task in industrial or household scenes. Motivated to learn this task autonomously and efficiently, we propose a goal-task-driven hierarchical sharing learning method (GTHSL). First, a hierarchical hybrid control (HHC) model is designed as the basic framework, which can reduce the difficulty of learning the push-grasp-place skill from scratch by decomposing this long-horizon skill into three position-control policies and an open–close policy. Then, a goal-task-driven shared policy network is designed to generalize these policies within the HHC to improve the learning efficiency and reduce the model size of skill. Furthermore, a multitask distributed training environment is built to facilitate the parallel learning of multiple policies in long-horizon tasks. The experimental results validate the effectiveness of the proposed method. The robot can successfully learn to pick and place designated targets in obstacle environments from scratch, with higher learning-efficiency and less model parameters.