Five Bridge projects ready to start

The benchmark and code in this project will be completely open-source, and it will be available to anyone who wants to modify it, providing a public sandbox that researchers can use based on their needs. The environments will resemble a subset of an actual video game in which different types of known bugs and common issues will be included. The goal of the project is to deliver lightweight and simplified versions of real game scenes to advance ATA and related systems.

Modern video games have grown in size and complexity and with that the need to test them. However, thorough coverage is often not feasible using manual human play testing. ATA systems have emerged as a remedy to this challenge but still fall short of human performance in finding bugs. ATA is far from solved, and as with any emerging technology, it is challenging to find environments that resemble modern games to be able to do research and test the algorithms. The standard benchmark environments that algorithms are evaluated on are often too simplistic and do not sufficiently resemble modern games. For this reason, researchers who want to do research in this field often need to construct their own environment for testing – or rely on benchmarks that are less relevant to modern games that are far more diverse and complex. Access to a more realistic environment will also be usable for other challenges, including complex 3D navigation, motion planning, predictions in dynamic environments, predictive models for finding bugs and issues, sample efficiency exploration, and other opportunities for testing and benchmarking AI. This would be valuable also to non-gaming related industries, including automotive, robotics, and transportation.

In the long term, the project participant aspire that the developed environment will become a part of the main benchmark suite used by both academia and industry in the evaluation of ATAs – and also to establish itself as a complement to the existing benchmark environments in reinforcement learning.

Participating researchers and industry partners

• ReaL – Linköping University: The Reasoning and Learning (ReaL) lab. Participants: Fredrik Heintz (Professor), Mattias Tiger (PostDoc), and Daniel de Leng (Research Engineer).
• SEED – Electronic Arts, currently a part of the WASP/WARA Media and Language. They also represent EA. Participants: Alessandro Sestini (Research Engineer, EA), Linus Gisslén (Technical Director, EA), Konrad Tollmar (Research Director, EA, Associate Professor, KTH).
• King – AI Center of Excellence at King. King has a WASP-affiliated industrial Ph.D. student working on the topic of applying generalization and scalability of reinforcement learning for game playtesting. Participants: Sahar Asadi (Director of AI Labs at King), Sara Karimi (Industrial PhD Student at KTH and King, AI/ML Engineer at King), Björn Brinne (Sr Director of AI Center of Excellence at King).

To address real-world challenges, ML researchers must have reliable access to large amounts of domain-specific data. At the same time, the industrial partners affiliated with WASP often possess rich and unique datasets that lack analogues among those publicly available. To promote collaborations between these parties, it is therefore desirable to provide a safe and effective infrastructure for data sharing. Since model evaluation is an integral part of ML research, this infrastructure should ideally include benchmarking services. Adding model hosting and generation of synthetic data from real-world data, when the latter is too sensitive to be shared directly, would lower the threshold for collaboration further.

The participant in this project aim to create a data management platform (DMP) that contains:

• a data directory
• automated tools for benchmarking and other forms of model evaluation
• API-access to trained models and generation services

A follow-up, national-scale, infrastructure project could be pursued as a joint venture together with Vinnova, and there are tentative discussions in this direction.

The project is lead by Anastasia Varava at SEBx/WARA M&L and Johanna Björklund at Umeå University.

In many unexpected traffic scenarios, automated vehicles will need to cooperate with connected infrastructure to address safety challenges such as dangerous occlusions or hard-to-predict pedestrian movement. To support automated vehicles in these scenarios, the connected infrastructure needs to help facilitate shared sensing, control, and decision-making across traffic. An important part of connected infrastructure are intelligent intersections. Intelligent Intersections are agents that enhance the awareness and coordination of vehicles passing through intersections. The primary objective of this project is to design and implement an intelligent intersection that can improve the safety of traffic by leveraging novel control algorithms, 5G networking technology, and tight integration with mature automated driving systems.

Specifically, in this project, WASP researchers, Ericsson, and Scania will design and implement an intelligent intersection on a 5G-equipped edge server at the Kista Innovation Park that will support shared situational awareness and safe coordination of connected and automated vehicles in intersections. While they validate recent research results from WASP researchers on intelligent intersections in a real 5G network, they will also focus on developing a platform at KIP for continued research into new intelligent intersection functionalities for enhancing the safety and efficiency of the transport network.

Participants:

KTH:
Jonas Mårtensson, Professor
Karl H. Johansson, Professor
Frank J. Jiang, WASP PhD Student
Xiao Chen, WASP Affiliated PhD Student
Kaj Arfvidsson, Research Engineer

Ericsson:
Anders Nohlgren
Hans-Christian Lindh Rengifo
Tomas Nylander
Henrik Aspegren

Scania:
Henrik Pettersson
Vandana Narri, Industrial WASP PhD Student
Natalie Richardson

Read more details at this project page.

Our online activities can reveal much about our thoughts, opinions, and interests. For these and other reasons, privacy-aware groups have long pushed for the use of encryption and anonymization techniques (e.g., HTTPS, VPNs, Tor, or iCloud Private Relay). Unfortunately, this is often not enough to protect users, as fingerprinting attacks and traffic analysis methods have been shown capable of extracting information about client activities and other sensitive information from encrypted network traces. The participant in this project want to accelerate the research within this field and to allow a greater focus to be placed on the implementation and evaluation of possible countermeasures to protect against such attacks. They point out that timely solutions for this are very important for the industry and are needed to protect against today’s attackers.

Fingerprinting defenses can be broadly categorized into four categories: imitation, regulation, alteration, and traffic splitting. Imitation aims to make network traces appear similar to others, regulation aims to make all traces look similar, and alteration attempts to modify the traces so as to remove any useful patterns. These three categories operate by adding fake data (padding) and/or modifying the packet timings by delaying individual packets. Finally, with traffic splitting, the traffic is split over several network paths, providing privacy under the assumption that an attacker can only observe one of the paths. In this industry bridge project, they will perform a comprehensive study of layered defenses, i.e., combination of several fingerprinting countermeasures. They will develop an evaluation framework and use it to objectively evaluate different layered countermeasures. The study will provide insights into which countermeasures work well together, which do not, and what it takes to provide good protection. When possible, they will use and build upon existing state-of-the-art frameworks.

The results would be highly relevant for the industry. For example, Sectra develops several high-end security solutions (e.g., VPNs) that could benefit from more efficient traffic analysis protection. The project also aligns well with the scope of WASP’s mission and profile, especially since cyber security recently has been highlighted as a high-priority research area within WASP.

Participants:

David Hasselquist, Johannes Wilson, Niklas Johansson, Mikael Asplund, Niklas Carlsson.
Linköping University, Sectra Communications.

The AD-EYE phase 2 builds upon the previously supported and recently concluded “AD- EYE” WASP industry bridge project, which was instrumental in providing engineering resources for the testbed development and in establishing new collaborations including with Volvo Cars. This phase 2 has the main goal to mature the testbed to an operational state and in obtaining a Transportstyrelsen (TS, the Swedish Transport Agency) approval for real road tests. The targeted Operational Design Domain (ODD) for the duration of the next 12 months involves low speed driving with a safety operator on two intersecting roads at the KTH campus Valhallvägen. KTH has the mandate to instrument these roads as needed with roadside sensors.

The AD-EYE testbed is developed and maintained by the Mechatronics division at KTH.

Participants

• Martin Törngren, KTH professor, WASP faculty
• Håkan Sivencrona, Zensaect
• Rafia Inam, Ericsson, Sector manager Trustworthy AI Research
• Stina Carlsson, Volvo Cars, Research project manager Safe Vehicle Automation
• Azra Habibovic, Scania – Technology leader automation
• Ahmed Terra, Ericsson, WASP industrial PhD student
• Magnus Gyllenhammar, Zenseact, WASP Industrial PhD student
• José Manuel Gaspar Sánchez, KTH, PhD student
• Jana Tumova, KTH Associate professor, WASP faculty
• Fredrik Asplund, KTH, Assistant professor
• Rusyadi Ramli, KTH, WASP affiliated PhD student
• Gianfilippo Fornaro, KTH, WASP affiliated PhD student
• Naveen Mohan, KTH, Affiliated WASP PhD student

https://www.adeye.se/