The paper introduces the Chat2Scenario framework, which utilizes large language models (LLMs) to enhance the process of extracting driving scenarios from naturalistic driving datasets. The key highlights are:
The framework employs the OpenAI GPT-4 LLM to interpret scenario descriptions in natural language, expanding the range of searchable scenario types beyond what was previously possible.
It provides a set of criticality metrics to quantitatively evaluate the extracted scenarios, allowing for more targeted selection of the most relevant scenarios for automated driving system (ADS) validation.
The framework is delivered as a user-friendly and shareable web application, making it accessible and practical for ADS testing engineers.
The paper first defines the key concepts of "Activity" and "Event" in the context of driving scenarios. It then outlines the dataset format used (highD dataset) and the overall methodology of the Chat2Scenario framework.
The scenario understanding module leverages prompt engineering to guide the LLM in accurately classifying vehicle activities (longitudinal and lateral) and their relative positions. The scenario searching module then evaluates the congruence between the LLM's responses and the vehicle trajectories in the dataset.
To promote the selection of the most relevant scenarios, the framework incorporates a criticality analysis module that computes various metrics, such as deceleration to safety time, required acceleration, and time to collision.
Finally, the framework generates the extracted scenarios in two widely used simulation formats: ASAM OpenSCENARIO and IPG CarMaker text. This enables the seamless integration of the extracted scenarios into simulation environments for ADS validation.
The paper presents qualitative and quantitative evaluations of the framework, demonstrating its effectiveness in extracting and reconstructing typical driving scenarios like following, cut-in, and cut-out. The quantitative analysis on a selected dataset file shows promising results in terms of accuracy, precision, recall, and F1 score for the identified scenarios.
toiselle kielelle
lähdeaineistosta
arxiv.org
Tärkeimmät oivallukset
by Yongqi Zhao,... klo arxiv.org 04-26-2024
https://arxiv.org/pdf/2404.16147.pdfSyvällisempiä Kysymyksiä