Domain Model Acquisition from Binary Traces

Abstract

Automated planning systems require symbolic models, but real-world data is often collected in subsymbolic formats such as binary encodings. We address this gap by introducing a domain model acquisition algorithm that handles subsymbolic state representations. Our algorithm takes as input plan trajectories where states are represented as binary vectors rather than fluent sets, along with action signatures, predicate definitions, and type hierarchies. It then simultaneously learns the mapping from binary state representations to symbolic fluents and constructs lifted action models with preconditions and effects. Across all evaluated benchmark domains from the International Planning Competition, our algorithm accurately maps bits to fluents and the learned action models closely match reference representations, despite having access only to binary encodings.