Predicate Invention for Bilevel Planning

Efficient planning in continuous state and action spaces is fundamentally hard, even when the transition model deterministic known. One way to alleviate this challenge perform bilevel with abstractions, where a high-level search for abstract plans used guide original space. Previous work has shown that abstractions form of symbolic predicates are hand-designed, operators samplers can be learned from demonstrations. In work, we propose an algorithm learning demonstrations, eliminating need manually specified abstractions. Our key idea learn by optimizing surrogate objective tractable but faithful our real efficient-planning objective. We use hill-climbing over predicate sets drawn grammar. Experimentally, show across four robotic environments able quickly solve held-out tasks, outperforming six baselines.