Can robots learn task rules by reasoning about examples?

Teaching robots from demonstrations usually captures how to execute a task, but not why. This work uses inductive logic programming to decompose complex tasks into interpretable symbolic rules at multiple abstraction levels. The system infers logical rules from demos and domain knowledge, then reuses them for higher-level reasoning. In block-assembly experiments, learned rules generalize strongly to harder tasks with novel objects, showing decomposed symbolic reasoning can achieve both transparency and generalization.