When both Grounding and not Grounding are Bad -- A Partially Grounded Encoding of Planning into SAT (Extended Version)

Researchers João Filipe and Gregor Behnke have introduced a novel approach to classical AI planning that addresses a fundamental trade-off in the field. Classical planning problems are typically defined using lifted first-order representations, which offer compactness but present computational challenges. Traditional methods either fully ground these representations (causing exponential size blowup) or operate entirely on the lifted level (avoiding grounding but complicating reasoning). The researchers propose a middle ground with three new SAT encodings that keep actions lifted while partially grounding predicates.

This hybrid approach delivers significant performance improvements over existing methods. Unlike previous SAT encodings that scale quadratically with plan length, the new method scales linearly, making it particularly effective for longer planning horizons. The empirical results demonstrate that their best encoding outperforms state-of-the-art approaches in length-optimal planning, especially on domains that are traditionally hard to ground. This represents an important advancement in automated planning efficiency and scalability.

The technical innovation lies in the partial grounding strategy, which selectively grounds only certain elements of the planning problem while maintaining the lifted structure of actions. This balanced approach avoids the worst aspects of both extremes: the computational explosion of full grounding and the complexity of fully lifted reasoning. The research was published on arXiv with the identifier 2603.19429 and falls under the Artificial Intelligence (cs.AI) category, representing work at the intersection of logic, symbolic computation, and planning algorithms.