GBAGC-RL: Goal-based arm-gripper coordination reinforcement learning approach for robotic manipulation skills

Recent research on robotic manipulation via reinforcement learning (RL) has garnered significant attention. However, RL faces hurdles in complex tasks because of high state–action dimensions and reward design complexities. It is important to find an easy-to-use framework to quickly achieve the representation, learning and generalization of robotic manipulation skills. This article proposes a novel manipulation learning method for complex robotic tasks. The key insight is that all complex manipulation tasks involve coordinated arm-gripper collaborative movements in the task space. By using a task representation and subgoal extraction algorithm to discern motion patterns and subgoals, this method addresses the “what to do” aspect of robotic manipulation tasks. Subsequently, it integrates goal-based hierarchical reinforcement learning (HRL) with pretrained foundational skills to address the challenge of “how to do”. This framework, called “goal-based arm-gripper coordination RL” (GBAGC-RL), integrates task representation, subgoal extraction, and goal-based hierarchical reinforcement learning to attain efficient and transferable robotic manipulation skills while drastically simplifying the design of the reward function. Simulation evaluations on multiple complex manipulation tasks demonstrate that the proposed framework exhibits strong generalization and transfer capabilities, outperforms many leading RL methods, and achieves higher task success rates with more stable manipulation skills. • A novel RL framework for robotic manipulation skills learning. • Speeds up training via task-type recognition and subgoal extraction from few demos. • Enables complex skills learning with simpler reward design. • Facilitate transfer learning of robot skills via task-type differentiation.