Fluid Imitation Discovering What to Imitate?

作者：Mohammad Yasser^*; Nishida Toyoaki

来源：International Journal of Social Robotics, 2012, 4(4): 369-382.

DOI：10.1007/s12369-012-0153-z

• Learning from demonstrations
• Imitation learning
• Fluid imitation
• Action segmentation
• Motif discovery

Research in learning from demonstrations has focused primarily on the question of how to utilize demonstrations to learn new behavior assuming that the demonstrator (teacher) explicitly teaches the learner. In this paper, we focus our attention on learning from unplanned demonstrations. In such cases, the learner has to take the initiative and decide for itself what actions of the-assumed unaware-demonstrator are to be learned. Moreover, the actions of the demonstrator are not pre-segmented and the learner needs to discover the boundaries of interesting behaviors without any help from the demonstrator. We propose a fluid imitation engine that augments a traditional LfD system. The proposed engine casts the problem as a well-defined constrained motif discovery problem subject to constraints that are driven from object and behavior saliency, as well as behavior relevance to the learner's goals and abilities. Relation between perceived behaviors of the demonstrator and changes in objects in the environment is quantified using a change-causality test that is shown to provide better results compared to traditional g-causality tests. The main advantage of the proposed system is that it can naturally combine information from all available sources including low-level saliency measures and high-level goal-driven relevance constraints. The paper also reports a series of experiments to evaluate the utility of the proposed engine in learning navigation tasks with increasing complexity using both simulated and real world robots.

• 出版日期2012-11