Rob-LAPI(S)²: Robotic Language-to-Action Planning via Iterative Schema Synthesis

Translating natural language (NL) instructions into formal symbolic plans is a fundamental challenge in AI planning. While Large Language Models (LLMs) offer language understanding, benchmarks show that 96% of LLM-generated PDDL parses correctly yet only 25% is semantically valid, revealing a gap between syntactic fluency and semantic correctness that external validators alone cannot close. We introduce Rob-LAPI(S)², a domain-agnostic pipeline that iteratively elicits and refines PDDL domain and problem files from NL, guided by symbolic validator diagnostics and static reachability analysis, without injecting any ground-truth structured knowledge. We evaluate Rob-LAPI(S)² on seven standard IPC planning domains, reaching 96% average symbolic validity and 73% ground-truth executability under simulator-grounded evaluation, a systematic validation gap that reveals the limits of formal self-consistency as a proxy for real executability. On household benchmarks (VirtualHome, AlfWorld) it achieves 90% and 55% task success from purely NL. Code: https://github.com/EmanueleMusumeci/Rob-LAPIS.