GALVATRON

Problem Context

Safety engineering is of the highest importance in the automotive industry. Self-driving trucks increasingly use data-driven AI to navigate complex situations, which poses unique challenges to system engineering. Two important challenges addressed in this project are specifying correct behaviour of the truck in a scalable, non-conflicting way; and to do failure source analysis for a perception system with unlabelled data.

Solution

We tackle the first challenge by moving towards model-based system engineering. Instead of text-based behaviour specification, we turn to simulation-based specification, using a context-aware motion planner. The behaviour is determined by a set of behaviour rules that activate depending on the context, thus offering a scalable, non-conflicting way of specifying behaviour. The second challenge is tackled by harnessing the power of Vision-Language Models (VLMs). Given an unlabelled dataset of images where the perception system failed, we use VLMs to extract tags about the environment depicted in these images, to reason about which tags caused the perception system to fail.

Results

For the first challenge, we have built a software tool that takes as input a scenario and outputs either the behaviour specification for that scenario or a simulation of the desired behaviour, based on reasoning about the desired behaviour in a knowledge graph. Currently this tool works for a limited set of scenarios, and we are working to extend this number to show the scalability of the method. For the second challenge, we developed a software tool that, given an input image, automatically outputs a limited set of tags that describe the environment of the self-driving truck, based on two public VLMs. We are working to extend the number of tags and improve the accuracy of the tool.