Thanks! The demo just shows the final agents after training (30K gradient updates). Interesting work re the reward maximizing curricula. I have not seen this before, so thanks for the pointer.
Yes, the so called "vertical response time" is still a factor to be considered, though I'd argue this stage is on equal footing as in the case of actual paramedics, assuming the Uber drivers are given a clear mandate to enter the environment in these scenarios.
Regarding your point about whether drivers could really become qualified, I see no reason why not. CPR/AED training exists for a reason, and it has saved many lives. There is also precedent of providing emergency AEDs for civilians to operate in these life or death cases. See the typical high school gym, which will likely sport an AED. Perhaps it wouldn't be every driver's cup of tea, but I can see many drivers sign up if well-designed incentives are put in place.
Will drivers confuse symptoms of SCA with stroke? That's why they would be required to go through AHA's training course.
You're right that AEDs can cost around $1000. Consider this in the context of how cars cost many times more than this amount. Further, the cost can be subsidized by Uber, as Uber stands to benefit financially when they save someone's life.
I agree with you that there is a lot of leg work required to make the further optimization of emergency lights/sirens. That would be something that comes later, and perhaps not necessary given that the median wait time is already under 3.5 minutes.