TLDR: AGENTS-LLM is a new framework that uses Large Language Models (LLMs) with an ‘agentic’ design to automatically generate challenging traffic scenarios for autonomous driving testing. It takes natural language instructions to augment real-world scenarios, offering fine-grained control and performing well even with smaller, cost-effective LLMs. Human expert evaluations confirm its ability to create high-quality, realistic scenarios that effectively challenge state-of-the-art autonomous driving planners.
Testing and evaluating autonomous driving systems is a monumental task, especially when it comes to rare but critical scenarios like dangerous or erratic driving. Traditionally, relying on real-world data means collecting massive datasets, which is expensive and ethically challenging for safety-critical events. While automatic scenario generation exists, data-driven models often lack fine-grained control and can introduce a ‘distributional shift’ if they create entirely new scenarios, undermining evaluation validity. Manual augmentation by human experts, though effective, simply doesn’t scale to the demands of modern self-driving system evaluation.
Addressing these limitations, a novel framework called AGENTS-LLM has been introduced. This system leverages Large Language Models (LLMs) in an ‘agentic’ design to augment real-world traffic scenarios using natural language descriptions. A key innovation of this approach is its agentic structure, which provides precise control over the generated output and allows for high performance even when using smaller, more cost-effective LLMs, rather than relying solely on large, expensive models.
The AGENTS-LLM framework operates with a Scenario Modifier Agent (SMA) that takes an original scenario and user instructions in natural language. The SMA then generates an updated scenario with the requested modifications. To ensure accuracy and adherence to user intent, the framework includes an optional Quality Assurance (QA) loop. This loop can be text-only, where a QA agent verifies the output against user instructions and common problems, or a hybrid visual-text approach, where a QA Engineer generates questions and a Visual QA Agent uses a rendered image of the modified scene to answer them, providing feedback for corrections.
To enable LLMs to process traffic scenarios, the system uses a text-based representation where traffic agents, lanes, and areas are described as vectors of attributes. The framework also allows for ‘function calls,’ enabling the LLM to retrieve precise coordinates along lanes for accurate placement of elements. This detailed control is crucial for creating specific safety-critical situations, such as placing a vehicle at an exact distance from an intersection to block a view.
Extensive human expert evaluations have demonstrated the framework’s ability to accurately follow user intent, generating high-quality augmented scenarios that are comparable to those created manually. The evaluation focused on two main aspects: placement accuracy and visual appearance. While placement accuracy measures how precisely traffic agents are positioned, visual appearance was assessed through pairwise comparisons by human domain experts, who rated which generated scenario they preferred in a blind comparison.
The results showed that while top-tier LLMs like GPT-4o performed exceptionally well even with simpler prompting, advanced strategies like ‘function calling’ (FC) and ‘Quality Assurance’ (QA) significantly improved the performance of smaller, more economical LLMs like Gemini-1.5-Flash and Llama3.1-70B. This indicates a promising path to reduce the reliance on expensive, closed-source commercial APIs. Notably, the visual QA variant allowed Gemini-1.5-Flash to achieve visual quality almost on par with the best models.
Also Read:
- GraphTrafficGPT: Advancing Traffic Management with Graph-Based AI
- Generative AI Enhances Realism in Human Motion Simulation
Furthermore, the generated scenarios proved to be highly challenging for state-of-the-art planning algorithms, such as the PDM-Closed planner in the nuPlan simulation framework. This suggests that AGENTS-LLM can effectively fuel research into more sophisticated planning methods by providing a scalable way to generate difficult test cases. The framework represents a significant step towards automating the creation of diverse and challenging traffic scenarios, crucial for the safety and generalization of autonomous driving systems. For more details, you can refer to the research paper.
