From my point of view these were the main aspects touched by the presentation:

1. Technical framework aspects they had to solve within ROS 2 that would have prevented certification. They call these the real-time gaps, see image below. This is mainly: runtime memory allocation, exception handling, real-time capable middleware, threading, and scheduling. Apex.AI fixed these issues by plugging in own versions of the allocator, a threading library, their own middleware, scheduler, etc.. Apex.AI Cert is also relying on a safe OS, e.g., QNX., see right-most column in the image above.
2. For 24(!) ROS 2 C++ packages that Apex.AI considers „safety-related“, they performed a hazard and risc analysis, wrote ~300 requirements, and performed FMEAs. According to Apex.AI, they spent roughly 14 person years to do so and to add the according tests and traceability to reach 100% MC/DC coverage (required by ISO 26262 for ASIL-D).
3. A tool qualification process for code generators, etc., that are for example part of the middleware layer.
4. For issues identified during the FMEAs that could not be mitigated on a code level, they provide a safety manual with usage restrictions that need to be adhered to when relying on the certification of Apex.OS Cert.

All of this took 5 iterations with TÜV Nord, resulting in roughly ~200 A4 pages submitted as a safety case. With that, Apex.AI provides a blueprint on how to certify an existing open-source community project for ASIL-D with the approval by TÜV Nord. This might be worth considering for application of further open-source software in safety-critical context.

While the process doesn’t seem to involve any magic, it saves Apex.AI’s customers roughly 14 person years in doing it on their own.

The post Apex.OS Cert – „How ROS 2 was Safety-Certified for Automotive“ appeared first on arne-nordmann.de.

CLIP is fooled by an adversarial patch, i.e. putting text in the images. Or as someone on twitter put it: better wear your „I am not robbing this bank“ t-shirt the next time you attempt to rob a bank.

More in the OpenAI blog: https://openai.com/blog/multimodal-neurons/

The post The latest and greatest in Adversarial Examples appeared first on arne-nordmann.de.