Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·Jun 9·5 min read
From USD Scenes to Knowledge Graphs: Zero-Shot Ontology Grounding with LLMs
TL;DR
Grounding objects in 3D simulation scenes to formal ontology classes is challenging due to reliance on brittle, manually curated dictionaries. This research explores using large language models (LLMs) to automate this grounding process for Universal Scene Description (USD) scenes without prior training.
✦ Why It Matters
Engineers can leverage LLMs for efficient ontology grounding in robotic applications, reducing manual effort and improving scalability.
Key Takeaways
- Explore LLMs for automating ontology grounding tasks.
- Reduce reliance on manual dictionaries in robotics.
- Test grounding methods on diverse 3D scenes.
- Implement findings to enhance robot task reasoning.
Full Summary
Grounding scene objects to formal ontology classes is crucial for enabling robots to reason about tasks in 3D environments. Traditional methods depend on manually curated dictionaries, which are often inflexible and do not generalize well.
This study investigates the use of large language models (LLMs) as a zero-shot, training-free method for automating the grounding process specifically for Universal Scene Description (USD) scenes. The researchers applied this technique to a kitchen scene containing 125 objects, leveraging the SOMA-HOME Ontology for classification.
Results indicated that LLMs could effectively identify and categorize objects without prior training, significantly reducing the need for manual intervention. This advancement suggests a more scalable approach to knowledge graph construction in robotics, potentially enhancing task reasoning capabilities.
The findings imply that LLMs can bridge the gap between raw scene data and structured knowledge representation.