A Multi-expert Agent for Efficient Learning from Demonstrations

In recent years, a myriad of seminal works on intelligent agents have been proposed, thanks to the advances in machine learning. Nonetheless, limited training efficiency and lack of transfer ability have been impeding the applications in broader domains, especially for human-robot collaboration when fast learning and flexibility are a must. To surmount this problem, we refer to a multi-expert architecture that contains pre-trained skills which can be transferred and reused easily. Meanwhile, a controller is employed to govern the learnt skills by unfolding them in a flexible sequence based on task specific preferences. This sequence is automatically interpreted from human demonstrations, minimizing the need for explicit procedural knowledge. To improve learning efficiency, we introduce reward augmented imitation learning and observation minimization techniques. Experimental results demonstrate that our algorithm is able to robustly pick up a multi-stage skill from only a few demonstrations, thus showcasing its capability for fast learning from human experts.