ABSTRACT
Model-based learning and model-free learning are both powerful approaches for teaching robots novel skills. Generally, they follow the same pipeline: during the training stage, the desired control policy is either constructed by modeling or regression from training data, which tries to capture the system’s causality and represent how the system progresses. The achieved policy is supposed to be universal and applicable. Therefore, it is applied during the test stage, i.e., new actions will be generated according to the test states and the trained control policy. However, for some complicated tasks, neither sophisticated models nor a large amount of training data can be easily obtained to train the control policy. This chapter proposes a novel pipeline for robot learning, named analogy learning, which is both model-free and data-efficient. This new pipeline focuses on finding correlation, instead of deriving causality. It focuses on finding a transformation between the training and test scenarios, instead of formulating a control policy to relate the training states and training actions. A robust non-rigid registration method, structure preserved registration (SPR) is introduced for implementing analogy learning. Analogy learning is also compared with traditional learning approaches to illustrate its advantages.
