Obsidian/.archive/Literature Notes/Notes on Papers/On Using Real-Time Reachability for the Safety Assurance of Machine Learning Controllers-Note.md

First Pass

Category: Experimental Results

Context: They used a F1/10 model of an autonomous car to test out a simplex safety structure on a neural network based controller. They try several different neural net types. Their trick is they use real time reachability to tell when to switch between an optimal controller and the simplex guard.

Correctness: They seem to do things pretty well and by the book. All of their explanations make sense and they do a good job citing sources. They do punt when it comes to talking about the formal verification of the switching mechanism, but they do make note that's future work.

Contributions: They show how a simplex system can work and the difficulties of the sim2real transition for machine learning controllers.

Clarity: Really nicely written.

Second Pass

What is the main thrust? They use a simplex style controller setup with real time reachability to know when to use a optimal ML based controller vs. a safety oriented controller. They use the reachability to do this in real time, and demonstrate how different ML models line up against one another.

What is the supporting evidence? They ran a whole bunch of experiments. They published all of the results, with their main metrics being Mean ML usage.

What are the key findings? The biggest findings are that when obstacles are introduced (or other general advesary behavior), the amount that the safety controller kicks in goes up by a lot. Sims can use around about 50+% ML controller based on the speed of the car, but when using the real track, the ML useage goes down below 20%. They also note their real track is much smaller than the simulation track.

Third Pass

Recreation Notes:

Hidden Findings:

Weak Points? Strong Points?