Explainable Robot Learning

Robot learning is a research field at the intersection of robotics and machine learning, which includes reinforcement learning and imitation learning etc. Deep imitation learning methods normally utilize a deep convolutional neural network (DCNN) with multiple layers to map pixels of visual observations to actions, which are trained via a supervised learning manner. However, traditional deep imitation learning has inherent black-box effects; that is, the decision-making process is not transparent to users. The black box effects can be acceptable in certain domains such as playing Atari games, but may impede the general acceptance of intelligent robots in tasks that involve humans or in critical applications. Due to the high cost of failure caused by wrong decisions, robots are expected to have the ability of reasoning failures in the case of executing wrong actions in the physical environment. To this end, we aim to develop intelligent robots that can generate explanations before the task execution, through which the user's trust in the robot's behavior can be built.