A synthetic data and generative A.I approach to verifying and validating A.I

点击翻译按钮获取中文摘要

This proposal is a **feasibility study** entailing the following activities:1. To build a consortium of industry and academic partners to evaluate the suitability of generalisable evaluation methods for Artificial Intelligence (AI) and Machine Learning (ML) approaches.2. To ensure the commitment of the parties to join a potential submission for the second phase.3. To propose the development of novel validation and verification approaches for trusted and responsible AI and ML.4. To produce a technical report outlining the approach to be developed in the second phase.The specific themes that we are targeting are **data pre-processing** and **evaluation**.Our approach is to use the Connected & Automate Vehicle (CAV) market as background and approach to the feasibility study. There is an ever-increasing need to ensure safety and trust for CAVs. It has been proven that it is impossible to ensure safety for CAVs through vehicle-level testing only.Additional challenges exist due to the unbounded conditions of the massively wide operational domain design (OBD); thus, relying exclusively on real-world testing to demonstrate safety across the full spectrum of scenarios that vehicles might encounter in deployment is an impossible proposition.Real-world testing of a vehicle's performance requires access to large test ranges with vast expanses of varied corner and edge case scenarios corresponding to the vehicle's expected OBD. Virtual scenario generation is required for verification and validation (V&V), but the generation of realistic and trustworthy scenarios efficiently remains an open question. A solution is needed that can provide thousands of scenarios and equivalent miles on public roads, which can significantly reduce the development costs and times.**Areas of focus**We intend to conduct a feasibility study to investigate a V&V solution through blended virtual and physical approaches to address the thousands of scenarios and equivalent miles on public roads by providing additional contextual and semantic information of the scenarios, which will significantly reduce the development costs and improve efficiency by ~500 times.**Innovations**Our project proposes the integration of two different approaches, procedural and generative, to the V&V challenges. Our approach will enable the generation of virtual evaluation scenarios to address corner and edge scenarios that are difficult to replicate in real-life.**Relevance**The initial project proposes a solution for developing a service that is generalisable for different Automated Driving Systems (ADS) by providing quick inferences and efficiently generating robust enriched contextual labelled datasets that are enriched with 3D information.