Refining discovered symbols with multi-step interaction experience

In our previous work, we showed how symbolic planning operators can be formed in the continuous perceptual space of a manipulator robot that explored the world with its single-step actions. In this paper, we extend our previous framework by enabling the robot to progressively update the previously learned concepts and rules in order to better deal with novel situations that appear during multi-step action executions. Our proposed system can infer categories of the novel objects based on previously learned rules, and form new object categories for these novel objects if their interaction characteristics and appearance do not match with the existing categories. Our system further learns probabilistic rules that predict the action effects and the next object states. There rules are automatically encoded in Planning Domain and Definition Language (PDDL), enabling use of powerful symbolic AI planners. Using this framework, our manipulator robot updated its reasoning skills from multi-step stack action executions. After learning, the robot was able to build stable towers in real world, exhibiting some interesting reasoning capabilities such as stacking larger objects before smaller ones, and predicting that cups remain insertable even with other objects inside.