Yasaman Golshan
RAG-Enhanced Webots World Generation with LLMs.
Rel. Alessio Burrello, Daniele Jahier Pagliari, Luca Benfenati. Politecnico di Torino, Corso di laurea magistrale in Data Science And Engineering, 2026
|
Preview |
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (3MB) | Preview |
Abstract
Large Language Models (LLMs) have demonstrated strong capabilities in code generation and semi-structured outputs such as JSON and XML, yet their ability to generate robot simulation environments from natural-language descriptions remains relatively unexplored. This thesis investigates the automatic generation of \texttt{.wbt} files for the Webots robotic simulator from textual scene descriptions, addressing both a practical robotics authoring problem and a broader question of structured and spatial reasoning. The task is challenging because Webots extends the relatively niche VRML97 format, whose syntax and scene representation are less familiar to current LLMs than those of mainstream programming languages. This can result in syntactic errors, invalid node types, and structurally or spatially incorrect scene representations.
To investigate this problem, two generation architectures are evaluated: a direct text-to-WBT pipeline and a two-stage, compiler-based pipeline that first generates an intermediate JSON representation and subsequently translates it into WBT
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
URI
![]() |
Modifica (riservato agli operatori) |
