Learning Multiple Robot Manipulation Tasks with Imperfect Demonstrations

Deep reinforcement learning (DRL) has been considered to be an effective method when dealing with robot control such as dexterous manipulation. With the increasing complexity of humanoid hand, control task is getting harder. Furthermore, to solve the drawbacks of DRL methods that may cause sample inefficient and slow exploration process, incorporating expert demonstrations have often been considered. Collecting demonstrations using deep learning methods will inevitably introduce errors, which forces the demonstrations to be imperfect. To solve that issue, we propose a weighting strategy ensuring that each demonstration will be used effectively and correctly, encouraging agent to be more efficient during learning. In the simulation experiments, with our weighting strategy and proposed unified reward function, our trained policies have achieved an average success rate over 85%, superior to the state of the art method in multiple manipulation tasks. We also validate that our method emerges more robust performance when using imperfect demonstrations with errors. The experimental results indicate that our method can be leveraged under different types of robotic environments.