Our proposal EASe - Evaluate Autonomous Systems with Foundation Models was approved to receive funding from SSF’s call for Industrial PhD projects.
We aim for devising a validated methodology to establish Foundation Models (FMs) as experienced co-workers to systematically evaluate and improve the safety of autonomous systems (AS) - during development and at operation. We will break down DevOps for engineering AS into its two constituent parts to systematically uncover application potential for FMs’ strengths:
In particular for Ops, we will focus on how to process onboard data streams with FMs under real-time constraints by identifying approaches to safe-guard and accelerate FMs. Complementary to the FMs’ strengths, we will systematically map and assess their limitations when applied to engineer safety-critical systems.
We expect to establish a methodology on how to responsibly integrate FMs as an experienced co-worker within DevOps for safety-critical AS during engineering (“Dev”) and for onboard application (“Ops”) to systematically assess and improve an AS’ safety in domains that face continuously growing amounts of data coupled with a need for continuous system evaluation.
Our proposal R-SAFE received funding from the Wallenberg Foundation (WASP).
Services such as robotaxis push the expectations towards automated driving systems (ADS) as they do not have a safety driver onboard. Yet, such ADS inevitably face complex real-world challenges that already resulted in tragic traffic accidents. Thus, constructively arguing for safety to cover all possible driving scenarios does not scale and hence, is effectively impossible. Yet, an ADS at SAE Level 5 needs to also maneuver in unknown situations that may turn out to be unsafe but systematic, quantitative methods to manage AI risks originating from that type “unknown/unsafe” are lacking. Runtime monitoring (RM) for complex systems such as ADS are considered to be an option because traditional methods from software engineering (SE), which are typically applied during system design and development, would not scale with the needs of an ADS.
We aim for devising runtime monitoring (RM) strategies to increase safety in ML-enabled systems. We plan to achieve this by addressing the theoretical challenge of systematically determining robust discriminant functions for RM, as well as the practical challenge of finding computationally efficient implementations that scale for industrial needs.
Our proposal Foundation Models for Time-Series Automotive Large Scale Data received funding from Vinnova, 2024-2026.
We aim at validating the potential of Foundation Models (FM) for analyzing and generating plausible time-series data. Such data enable researchers to perform experiments, simulations and scenario analysis without relying solely on limited real-world data. We envision that our results will pave the way for a whole new generation of digital twins that combine real data, synthetic data, and FMs. Such next generation of Digital Twins may provide insights on the performance of a system in terms of accuracy and computational efficiency of the models in the context of traffic analysis.
Our proposal SACCADE received funding from the Swedish Research Council (VR) call for research projects within natural and engineering sciences.
AI/ML-enabled systems need large amounts of high quality data for training and evaluation. At industrial scale, such datasets are already referred to as data lakes to denote their size. Growing them by just adding more and more data drops is rather simple. However, determining the added value of the next drop to be inserted to a data lake has become increasingly complicated.
In SACCADE, we aim at investigating, developing, and systematically evaluating an automated approach to estimate the added value of new data to be inserted to a large-scale data lake in a computationally efficient way. We will exploit space-filling curves (SFC) to map multi-dimensional data as well as unstructured large multimodal data to their corresponding, single dimensional representations. For the latter, we will embody a deep learning (DL) approach to extract semantic-aware key features as input to SFCs.
We will use patterns emerging on SFCs to discover already present data samples and also to identify characteristics of still missing data samples in a data lake. Thereby, we expect that it will be possible to determine the level of data diversity and quality in existing data lakes to tackle important questions such as “How good are our datasets and when do we have enough data?”
Our proposal SAICOM from the team KTH, Chalmers, and GU received funding from SSF’s Future Software Systems call.
Already in our today’s society, we see a rapidly growing amount of distributed embedded systems that visisbly and invisibly support our daily lives and infrastructure. The next generation of wireless networks (5G and beyond) will allow to transfer larger data volumes to in shorter times. This will even further accelerate the permeation of such units that in a software-enabled way support infrastructure and processes like self-driving vehicles for example. We have also seen the growing importance of adopting AI/ML for such units to better cope with highly unstructured data and to find patterns therein for monitoring or control. This growing amount of distributed embedded systems and larger data volumes to handle will inevitably require a different approach to better deal with the engineering processes for such AI/ML-powered, software-driven embedded systems. In SAICOM, we will consider the wireless networks, their infrastructure such as base stations, distributed embedded systems, and the engineering processes for all these aspects in symbiosis to conduct research in what way AI/ML and highly distributed software can be fundamentally rethought.
Our proposal ASSERTED from the team Volvo Cars and Chalmers received funding from Vinnova call.
Software for self-driving vehicles needs to be designed and maintained in a way that allows to be improved based on findings and insights from continuous field observations–however, in a safe, secure, and traceable manner. The research goal this Vinnova FFI-funded research project is to explore methods for coping better with safety of autonomous driving in continuous development and deployment processes. The goal is to understand how the safety argumentation of autonomous driving needs to be adapted to this new context and what methodological and technical solutions can tackle the identified challenges at the scale of an automotive OEM.
Our proposal SHAPE-IT from the team Chalmers, GU, TU Delft, University of Leeds, University of Ulm, and TU Munich received funding from European Union’s Horizon 2020 under the Marie Skłodowska-Curie grant agreement 860410.
The overall goal of SHAPE-IT is to enable rapid and reliable development of safe and user-centered automated vehicles for urban environments. The main objective of SHAPE-IT is to facilitate the safe, acceptable integration of user-centered and transparent automated vehicles into tomorrow’s mixed urban traffic environments, using both existing and new research methods, designing advanced interfaces and control strategies. As AI is a core technology for automated vehicle development, this EU Horizon 2020-funded research network aims to integrate the knowledge of human factors with that of AI in automated vehicle development, reducing the gap between human-factors and AI scientists and automated vehicle software developers. The main educational aim of this project is to deliver a future generation of human factors researchers with an excellent multidisciplinary (cognitive and behavioural psychology, human factors, computer science, and engineering) education in human factors experimental and modelling methods, human-AV interaction, safety analysis and AV design.
Our proposal EVIDENT from the team AstaZero, GU, Chalmers, RISE, VTI, Einride, Asymptotic, and Veoneer received funding from Vinnova call.
Validation and verification activities for automated driving is using a mix of various ingredients such as closed-loop simulations for model-in-the-loop (MIL), software-in-the-loop (SIL), processor-in-the-loop (PIL), or hardware-in-the-loop (HIL), open-loop approaches based on replaying previously recorded sensor and vehicle data, systematic experimentation on proving grounds, and field monitoring on prototypical vehicles; the latter may even be realized as part of controlled A/B experimentation with inert software that is running next to production software. Defining the right mix of all aforementioned approaches is an engineering challenge, while the systematic quantification of the “fidelity gap” between pure virtual testing in simulation and controlled experiments on proving grounds with the real sensors and the real vehicle is still a challenge from the theoretical perspective for academia. This project will address the “fidelty gap” challenge to identify methodogical approaches to guide how to better quantifiably transfer and interpret results from pure virtual testing for real settings.
Our proposal ASPECT from the team Volvo Autonomous Solutions, GU, BTH, Ericsson, Telia, and Voysys received funding from Vinnova call.
The project ASPECT aims to design a system of electrified machines in a confined area. Together with partners in the project, we combine automotive and digital technologies to enable digitalized and energy optimized solutions for a site. Furthermore, the target is to create a testbed for a digital infrastructure for the system, a method to identify and evaluate requirements of the infrastructure for electrified machines, and virtual models for energy optimization.