RISK ASSESSMENT IN VLA MODELS
Full title: Risk assessment in vision-language-action models for autonomous driving
This project studies risk assessment in vision-language-action models (VLAs) used for automated driving (AD). Modern AD systems rely on deep learning for both perception and planning, and a major challenge is ensuring that they can measure their own uncertainty. In traffic there will be situations where the system is uncertain, for example due to unusual objects or behaviours, bad weather or blockages. If measured, this uncertainty can be used to adjust driving behaviour, for example by slowing down. The team has recently developed a system that monitors end-to-end AI systems for AD and continuously assesses collision risk. VLAs are trained on large amounts of text and show a rich understanding of the world, which is believed to help AD systems adapt to new and rare scenarios. VLAs can also use language for interaction: verbalising their situational awareness, answering questions and receiving text instructions. The project will test different strategies for extracting uncertainty from VLAs, including adapting the previously developed system for end-to-end models, and will identify use cases of uncertainty in automated driving through structured interviews with experts in the field.
Traffic safety benefit: Lets automated vehicles measure their own uncertainty and collision risk in rare or unusual situations, so they can adapt (e.g. slow down) before an incident occurs.
Connection to other projects: Connected to the ELLIIT project: F4:A Robust and Reliable Vision-Language-Action Interface
Keywords: vision-language-action models, automated driving, uncertainty estimation, risk assessment, deep learning, end-to-end driving