Application Story
AI-powered ADAS scenario generation and management
AI-powered ADAS scenario generation and management
Summary
Clients in the automotive industry face the challenge of creating and testing an infinite number of scenarios with limited resources and time, a problem that exaggerates if they are using different standards and simulation platforms. IAV Automotive Engineering leverages cutting-edge AI technologyto streamline the creation and management of ASAM standard-based scenarios and overcomes the constraints of traditional scenario generation.
By integrating customer-facing Large Language Models (LLMs) with rigorous data processing procedures, we transform diverse test case specifications and scenario descriptions into highly modifiable and parameterizable scenarios that are compatible with clients’ simulation and testing toolchains.
This automated approach, supported by ASAM standards, has significantly enhanced our clients’ efficiency, reduced costs, and improved the quality and flexibility of scenario management and test automation processes.
Initial Situatíon
Before the project, our clients faced significant challenges in the manual creation of scenarios, which was both time-consuming and prone to human error. This complexity was further compounded when the clients’ input and output requirements pole apart. The input sources may range from expert knowledge, crash reports, regulations, to internal function requirements. In addition, while companies need realistic scene setups for ADAS function validation, the input map might originate from multiple sources. The output formats supported by different simulation software were also incompatible, thus hampering our ability to quickly respond to new requirements and feed the output forward to testing processes. Therefore, the motivation was clear: we need a more efficient, accurate, and compatible solution.
Solution
Understanding the need for transformation, we integrated the text understanding capability of Large Language Models (LLMs) to automate the generation of these scenarios. The LLMs, combined with a pre-processing and post-processing techniques, enabled us to convert various test case specifications or text descriptions into scenarios that could be easily modified and derived. Moreover, this new solution harmonized with our internal scenario database management and analysistool chain. Specifically, users can perform scenario generation, select the algorithms for finding critical scenarios,conduct scenario-based simulations to validate ADAS functions and carry out SOTIF analyses. Byreferencing the ASAM OpenLABEL standard, we ensured that our scenario database store relevant scenarios in a highly organized and generic format, with data collection methods, assets, parameters, tags and KPIs carefully defined to fulfill clients’ need.
Challenges during the project
Throughout the project, we encountered several challenges, notably the integration of sophisticated AI with our existing tools and aligning it with ASAM standards.
Ensuring accuracy and reliability of the scenario generation pipeline posed a significant hurdle, requiring continuous refinement of our data processing steps. We tested multiple data sources and scenario categories to better hone the multimodal AI service for understanding and tagging different inputs. We conducted extensive research and statistical testing to ensure that LLM service is deployed in a safe, economic and stable way. We were transparent with users about how their data would be protected and provided options for deploying the pipeline in their internal environment. By using access control and authentication protocols, we verified that only allowed users could access the database and upload or download scenario data.
At the same time, since clients need scenarios during different phases of the V-model, we faced challenges when the scenarios became domain specific, necessitating precise filtering and organization. As a result, we studied how ASAM OpenSCENARIO, ASAM OpenDRIVE and ASAM OpenLABEL components can be used and combined to adapt to regulation scenarios, accident scenarios, and the concrete scenarios designed for functional purposes includingautonomous driving, parking and guard. We found out that the ASAM OpenX standards could be applied throughout the validation and verification of autonomous driving systems. Therefore, this standard had the potential to serve as a bridging language across different sections of ADAS development in the long term.
Moreover, we iterated on the post-processing and visualization components. We studied possible use cases and designed the UI/UX layer for users to interact with the AI tool in a clear and intuitive manner. We also ensured that the tool supports common types of simulation software integration, both locally and via cloud deployment. IAV engineers collaborated extensively during the releasing, productionizing and monitoring phase to seamlessly embed the AI tool into target systems.
Business Benefits
The transition to this innovative, AI-driven approach brought substantial business benefits. We witnessed time and cost savings by significantly reducing the time required to manually generate and organize scenarios. Additionally, the new process offered a gain in flexibility, allowing us to quickly adapt scenarios to new data sources, requirements and parameter variations. Most importantly, the quality and compatibility of our scenarios are guaranteed, ensuring outcomes that can be further built upon in mainstream simulation software and validation tools.
The ASAM standards have provided us with a robust foundation for scenario generation and tool integration, which is crucial for advancing our work in ADAS Validation. This alignment with industry standards not only streamlines our processes but also ensures that we remain at the forefront of technological innovation.
Mr. Tianyi Lian / Automated Driving System Department Head
Conclusion
By leveraging LLM and ASAM standards, we not only addressed our customers’ immediate challenges but also enhanced the effectiveness and scalability of our ADAS simulation tool chain. We believe that embracing AI technology would position ourselves for long-term success, maintaining a high standard of industry excellence.
BACK TO OVERVIEW
Follow us on
