Task-oriented hybrid learning for robotic skill acquisition and control

Accurate execution of metrological applications and complex manipulation tasks remains a significant challenge in smart manufacturing systems. Although tele-operated and pre-programmed robots improve efficiency and safety, their dependence on expert control and limited adaptability in dynamic environments restrict their deployment. To address this, we propose a task-oriented hybrid hierarchical learning framework to facilitate robot skill acquisition. This framework integrates imitation rewards from expert demonstrations with task rewards from environmental interactions, enabling more purposeful exploration and improved learning efficiency. It combines Behavioural Cloning (BC), Generative Adversarial Imitation Learning (GAIL), Proximal Policy Optimization (PPO), and Virtual Reality (VR) technology. Training involves two phases: first, VR expert demonstrations are used with BC to establish an initial policy; second, this policy is refined using PPO, incorporating both external rewards from the environment and internal rewards from the GAIL discriminator. Numerical experimental evaluations on three manipulation tasks demonstrate our method’s effectiveness, achieving high success rates and robust performance in challenging scenarios.