This years SCSSS was run as an one day workshop within the large SafeComp conference.

Below follows a summary with some personal highlights from the presentations two keynotes and six presentations.

Please check the conference web for access to all presentations

The workshop kicked off with the keynote, “Assuring Safety in the Face of the Unpredictable” by Mario Trapp, Fraunhofer. A main concept was that you had to take into account a triad of aspects when working with safety for new AI systems:

Towards the end of his talk he looked at the potential analogy of a set of intelligent systems interacting with each other and with humans to a society, where the systems could be governed by rules and laws similar to our current society.

William Zeng, from Roben with his presentation “Automotive Overview on System Safety Standards and Practices for ICVs and NEVs in China: A different approach”  gave a very interesting insight into the fantastic dynamism of the Chinese car industry, how they had in a very short time become a major driver in the industry. It is well worth to look a the detailed examples presented in the slides. Here is just one example of Xiaomi entry into the car market:

Michael Wagner from ECR.AI “Embracing Change: LLM Use in Safety Engineering” first gave an overview of possibilities and pitfalls with LLM models:

Then he looked at the possibility to use formal proof checkers to check the output of LLM models.

Lunch was served together with all the other participants from SafeComp.

The afternoon started with a tutorial by Hans Liwång, KTH/Försvarshögskolan on “Interaction between technical and social system for the defense and security”. Hans started with observations about the importance of training and preparedness for survivability in naval warfare based on the Falkland conflict. The military safety is always a balance between safety and being able to continue to operate and achieve the target. There is no safe state as for a passenger car when standing still.

He also pointed out the value of giving people the ability to utilize the systems outside the safety limits in emergency situations, which might run contrary to how we think about safety to always keep the system within safe operational conditions.

AI Safety Assurance in the Automotive Domain – Standards Mapping and Application for an AI-Based SoC Estimation Function by Fredrik Warg; RISE gave an overview of how AI and safety research and standards have exploded in the last years.

Ali Nouri, Volvo Cars talked about “AI-enabled DevSafeOps for Autonomous Driving Software”. I was specially impressed by the breath of Ali’s research into how AI would impact all aspect of safety development. It will be interesting to follow this research in the future.

Filip Strand, KTH talked about his PH.D. work on Countering wildfire risk with AI-enabled sensor platforms so that the machines could avoid stones etc. that could lead to sparks causing fires. One major problem here was to create enough training data of relevant terrain that is labeled.

One idea that came up in the discussion was to rather use a program to generate images to train on. This should be possible as there is a limited number of elements in the terrain, however they have infinite many shapes and are randomly distributed. The generator could generate a labeled image that the IA could train on and as it is already labeled there is no “human” interaction.

Adapting ISO 21448 SOTIF for Mobile Machinery: A Forestry Automation Case Study by Aria Mirzai, RISE looked at applying SOTIF to  a Shuttle transporting logs from a harvester along pre-mapped forest roads to drop-off point.

The workshop ended with a buffet receptions at the Stockholm Townhall sponsored for SafeComp by Stockholm City. Here our foreign guests could try out most Swedish specialties; pickled herrings, meat balls, graved salmon and many more.