4 ms·
Sounds like a great reason to train the model on a closed road with a trained (and consenting) driver playing the role of "lane-blocking-driver" (in that exampl
by cjrp 5y ago
Sounds like a great reason to train the model on a closed road with a trained (and consenting) driver playing the role of "lane-blocking-driver" (in that example).
- natch 5y agoBut that would take about a billion years, with all the variations and edge cases.
- RandallBrown 5y agoIt would take the same amount of time if they're on a closed course or in a city. The difference is you'd need to pay a lot more people to work on it.
- natch 5y agoOr you could have a large fleet of a million plus instrumented vehicles gathering data as they drive, then you could use that real world data to analyze and learn from all sorts of situations.
- _ea1k 5y agoYeah, and the real world has advantages as the response of human drivers is likely more realistic. I'd bet they use this data to feed their simulators.
- jonas21 5y agoThey also train in simulation, where they modify real scenarios to run through variations and edge cases. [1] As of last year, they'd driven 15 billion miles in simulation - which is more than a typical person will drive in a million years. Still a ways to go to get to a billion though. [1] https://blog.waymo.com/2020/04/off-road-but-not-offline--simulation27.html https://blog.waymo.com/2020/04/off-road-but-not-offline--sim...
- wolverine876 5y agoInteresting. What is the validity of the simulations? Are they randomized? Based on actual driving events?
- ahahahahah 5y agoThis isn't Tesla we're talking about, Waymo has a large testing facility where they do exactly that sort of thing. Here's an indepth report on it from 4 years ago https://www.theatlantic.com/technology/archive/2017/08/inside-waymos-secret-testing-and-simulation-facilities/537648/ https://www.theatlantic.com/technology/archive/2017/08/insid...
- cjrp 5y agoInteresting, thank you