Large Language Models (LLMs) are demonstrating outstanding potential for tasks such as text generation, summarization, and classification. Given that such models are trained on a humongous amount of online knowledge, we hypothesize that LLMs can assess whether driving scenarios generated by autonomous driving testing techniques are realistic, i.e., being aligned with real-world driving conditions. To test this hypothesis, we conducted an empirical evaluation to assess whether LLMs are effective and robust in performing the task. This reality check is an important step towards devising LLM-based autonomous driving testing techniques. For our empirical evaluation, we selected 64 realistic scenarios from DeepScenario--an open driving scenario dataset. Next, by introducing minor changes to them, we created 512 additional realistic scenarios while keeping them realistic, to form an overall dataset of 576 scenarios. With this dataset, we evaluated three LLMs (GPT-3.5, Llama2-13B and Mistral-7B) to assess their robustness in assessing the realism of driving scenarios. Our results demonstrate that: (1) Overall, GPT-3.5 achieved the highest robustness compared to Llama2-13B and Mistral-7B, consistently throughout almost all scenarios, roads, and weather conditions; (2) Mistral-7B performed the worst consistently; (3) Llama2-13B achieved good results under certain conditions but not for the others; and (4) roads and weather conditions do influence the robustness of the LLMs.
Python 3.8 or higher
pip install -r requirements.txt
The folder deepscenario contains all scenario files involved.
mutate_scenarios.py contains functions related to scenario mutation.
llm_api.py is used to generate prompts and call LLM API to output the answers.
In this script, api_key needs to be added manually:
client = OpenAI(api_key="")
...
fireworks.client.api_key = ""
api_key for OpenAI: OpenAI
api_key for fireworks.ai: fireworks.ai
parse_results.py involves all program implementations associated with the analysis of results.
The folder outputs_results contains all LLM output and analysis result files.