arXiv cs.AI· Bastien Dussard (LAAS-RIS, LAAS), Aur{\'e}lie Clodic (LAAS-IDEA, LAAS-RIS, LAAS), Guillaume Sarthou (LAAS-RIS, EPE UT, LAAS)·· 2 小时前AI 评分30
URDF2Ontology：面向机器人运动学的分类学无关神经符号化方法
Taxonomy-Agnostic Neurosymbolic Grounding of Robot Kinematics
AI 导读
研究者提出 URDF2Ontology，一个自动化的分类学无关神经符号框架，将原始 URDF 文件转化为填充好的 OWL 本体。该框架利用大语言模型（LLM）做针对性语义推理，配合迭代式图清除机制防止上下文窗口饱和，并采用感知形态的双轨架构：自底向上的 Tree Track 处理分支装配、自顶向下的 Chain Track 处理串联肢体。
正文 · 原文
Abstract:While symbolic reasoning for task allocation requires grounding a robot's physical capabilities, standard Unified Robot Description Format (URDF) files provide only low-level kinematics without the formal semantic context needed for reasoning. We present URDF2Ontology, an automated, taxonomy-agnostic neurosymbolic framework that transforms raw URDF files into populated, grounded OWL ontologies. It leverages a Large Language Model (LLM) for targeted semantic reasoning, coupled with an iterative sequential graph-purging mechanism to prevent context window saturation. To process complex kinematic structures, it employs a morphology-aware dual-track architecture featuring a bottom-up Tree Track for branching assemblies and a top-down Chain Track for serial limbs. Guided by neurosymbolic guardrails, the framework extracts functional sub-graphs and asserts populated ontologies aligned with arbitrary user-provided taxonomies. Empirical evaluation across a diverse fleet of heterogeneous robots demonstrates strong schema alignment, valid structural extraction, and a significant reduction in manual domain authoring overhead.
| Subjects: | Artificial Intelligence (cs.AI); Robotics (cs.RO) |
| Cite as: | arXiv:2611.13850 [cs.AI] |
| (or arXiv:2611.13850v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2611.13850 arXiv-issued DOI via DataCite (pending registration) |
|
| Journal reference: | 4th AAAI Fall Symposium on Unifying Representations for Robot Application Development, Nov 2026, Arlington, VA, United States. pp.895-903 |
Submission history
From: Bastien Dussard [view email] [via CCSD proxy]
[v1]
Tue, 10 Nov 2026 10:11:30 UTC (2,144 KB)
来源：arXiv cs.AI · arxiv.org