Control semántico de un robot rehabilitador mediante modelos de lenguaje

Abstract

Large Language Models (LLMs) offer new possibilities for simplifying interaction with robotic rehabilitation platforms, whose low-level parameterisation represents a recognised barrier to clinical usability. This work proposes an architecture for the semantic control of the Discover2Walk platform, a cable-driven robotic system for paediatric gait rehabilitation. An LLM-based agentic layer interprets natural language commands issued by the clinician and generates a structured action plan. A deterministic safety layer algorithmically validates each action against a catalogue of admissible ranges prior to execution. Finally, an actuation layer materialises the validated commands as ROS~2 messages to the platform modules. The architecture is validated in simulation under three experimental conditions and preliminarily on the real robot, demonstrating its preliminary viability to translate clinical intent into safe commands on the robotic hardware.