This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
AGENTS-LLM: Augmentative GENeration of Challenging Traffic Scenarios with an Agentic LLM Framework
Created by
Haebom
Author
Yu Yao, Salil Bhatnagar, Markus Mazzola, Vasileios Belagiannis, Igor Gilitschenski, Luigi Palmieri, Simon Razniewski, Marcel Hallgarten
Outline
This paper presents a novel framework for effectively generating rare but important scenarios for autonomous driving system evaluation. Existing data-driven models have problems such as requiring massive training data, difficulty in fine-grained control, and low evaluation validity due to distribution differences with existing data when generating new scenarios. In this paper, we propose an LLM-based agent framework that augments real traffic scenarios using natural language explanations. Through agent-based design, we maintain fine-grained control and high performance even when using small LLMs, and enable expert-level scenario augmentation.
Takeaways, Limitations
•
Takeaways:
◦
We propose a scenario augmentation framework based on LLM using natural language to improve the efficiency of autonomous driving system evaluation.
◦
We demonstrate that agent-based design enables high-quality scenario augmentation even with small LLMs.
◦
We verified that the scenario generation performance was comparable to expert-level manual augmentation.
•
Limitations:
◦
It may depend on the performance of LLM. The limitations of LLM may affect the quality of scenario generation.
◦
The quality of the generated scenarios can vary depending on the quality of the natural language description. Clear and detailed descriptions are required.
◦
It cannot guarantee perfect matching with the actual road environment. Further research is needed on the generalization performance of the model.
