Taming the LLM: Reliable Task Planning for Robotics Using Parsing and Grounding Chapter in Scopus uri icon

abstract

  • Domestic service robots require the ability to translate ambiguous human commands into reliable, executable action plans. While Large Language Models (LLMs) offer powerful natural language understanding, they are prone to hallucinating unfeasible actions, creating a gap between abstract command interpretation and physically grounded execution. This paper presents a multi-stage pipeline designed to bridge this gap for robust long-horizon task planning for robotics, making three primary contributions: 1) An automated process to generate a task-specific dataset in robot command interpretation for fine-tuning LLMs. 2) The implementation of a model-agnostic schema-aligned parsing system that guarantees syntactically valid plans and prevents action-level hallucinations. 3) A pre-execution grounding stage using a vectorized knowledge base to align commands with the robot¿s real-world perception. The system was scoped for and benchmarked using commands from the RoboCup at Home GPSR task. Experiments demonstrate that the fine-tuned model, deployed on the robot¿s embedded hardware (Jetson AGX Orin), achieves a 96.5% success rate on linguistically diverse commands, outperforming both the non-fine-tuned baseline (29.6%) and a leading cloud-based model (90.4%). Furthermore, the schema-aligned parser proved to be 11% more dependable than standard token constrained generation methods. The code, datasets, benchmarks, videos and website are available at: github.com/RoBorregos/frida-cortex, providing an accessible framework for developing reliable, real-world command interpretation on domestic service robots. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.

publication date

  • January 1, 2026