Generate-then-Rank: Learning Symbolic Goal Interpretation for Embodied Agents

This paper introduces a two-stage 'generate-then-rank' approach for goal interpretation in embodied agents, designed to translate natural language instructions into formal symbolic goals executable by planners. The method first uses supervised fine-tuning to adapt a Large Language Model (LLM) to the Embodied Agent Interface (EAI) output convention, followed by a Candidate-Output Ranking from Errors (CORE) module that learns preferences over generated alternatives. Evaluated on household scenes within the EAI Benchmark using multiple open-source LLMs including DeepSeek-V3 chat, DeepSeek-Reasoner, and Qwen2.5 variants, the proposed method outperforms baseline approaches such as simple prompting, chain-of-thought reasoning, and knowledge graph-based methods.