LLMs auto-populate robot ontologies from low-level URDF files
Robots get semantic understanding without manual ontology engineering.
A team from LAAS-CNRS (Bastien Dussard, Guillaume Sarthou) has introduced a new method that lets large language models automatically infer semantic abstractions for robots from their standard URDF files.
Their pipeline works by taking a robot's Unified Robot Description Format (URDF) – which typically contains only structural and kinematic data – and prompting an LLM with concepts from an existing ontology. The LLM then assigns meaningful labels to joints and links (e.g., identifying a rotational joint as an 'elbow'). To improve output reliability, the system combines majority voting across multiple LLM queries with syntactic and schema-level validation, ensuring the generated ontology respects formal constraints.
Validated on several robot descriptions, the approach demonstrates that LLMs can effectively bridge the gap between low-level mechanical descriptions and the grounded, explainable knowledge representations needed for human-robot collaboration. The work was accepted at the 18th International Conference on Social Robotics (ICSR 2026) in London.
- Pipeline transforms bare URDF structural data into semantically rich robot ontologies
- Uses majority voting over multiple LLM queries plus syntactic/schema validation to reduce errors
- Tested on multiple robot models, successfully inferring joint/link semantics for explainable reasoning
Why It Matters
Automates a critical bottleneck in cognitive robotics, enabling scalable, semantically grounded robot knowledge without manual ontology engineering.