FAST-PLAN: FAilure Driven Training Data Synthesis for Ambiguity-Robust Embodied Task PLANning in Open Source LLMs

Open-source large language models (LLMs) are increasingly deployed for embodied task planning, yet they remain brittle under ambiguous, human-like instructions characterized by implicit pronouns, underspecified attributes, and noisy dialogue. Consequently, current state-of-the-art open models often suffer severe performance collapse in realistic, ambiguous scenarios. To address this limitation, we introduce FAST-PLAN, a scalable, failure-driven data synthesis framework that trains robust planners using their own mistakes. Operating within the AI2-THOR environment, FAST-PLAN programmatically injects linguistic ambiguity into deterministic symbolic tasks. It employs an automated pipeline to generate vague instructions via an LLM augmentor, validate groundings, evaluate execution, and leverage a critic LLM to classify planning failures and generate refined feedback. We finetuned Qwen and Llama family models on synthesized trajectories generated by FAST-PLAN. The resulting models outperform baselines on REI-Bench by +50% on success rate. Furthermore, they exhibit strong zero-shot generalization on out-of-distribution ambiguity benchmarks, improving macro-F1 scores by up to 6% on the CLARA benchmark, highlighting the effectiveness of FAST-PLAN.