-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Waymo
- Thread starter Daniel in SD
- Start date
They've done cold weather testing in different parts of Michigan and New York state for several years. I think Buffalo is new though.
EVNow
Well-Known Member
How do they get a large amount of data... Isn't one or two cars insufficient?
diplomat33
Average guy who loves autonomous vehicles
Where did you get it was only 1-2 cars? The blog just says "small fleet". It is likely more than 1-2 cars.
EVNow
Well-Known Member
It probably is for data points specifically related to weather. They'll need enough data to create a skilful model of traction, vision issues, sensor reliability, etc. for simulations, but more cars in a single region will not add to that does not increase the variety of data (different weather patterns). It's the same data from additional sources which is less useful beyond a threshold.
Experiencing those 1 in 10 million mile events in the real-world in each type of weather condition isn't necessary or even desired. If you cannot rely on simulations for Quality Assurance then every single software change would require millions of miles in every weather condition prior to deployment. They need enough data to create a model, and prove the model skilful. Waymo vehicles already detect insufficient data issues and pull over during them (very heavy rain for example).
Friction (and thus stopping distance for a given speed) change but all the behaviours are otherwise the same and still apply. Simulations of random drivers braking too late remain similar but with larger distances and lower speeds applied.
Edit: Weather modelling is a component of automated train signalling systems too. Rain, ice, fallen leaves, abnormally high temperatures, etc. all impact stopping distance and thus the safe distance between trains in a moving block system. They don't experience every single level of rain, but create a regression curve from a few examples and use that.
Last edited:
diplomat33
Average guy who loves autonomous vehicles
Yes, I think 10 cars would be sufficient.
1) Waymo car collects far more data than a Tesla since it has more sensors. Each Waymo car is collecting high resolution video from dozens of cameras, high resolution point clouds from multiple lidar, short range and long range, as well as high resolution point clouds from multiple radars both short range and long range. And Waymo is collecting data from cameras, lidar and radar so they are getting much richer data than just images or video that Tesla would get.
2) Waymo can also use that real world data to generate even more synthetic data. So Waymo will actually get more data for training than just what they collected from the 10 cars driving around.
3) Quality is key. Waymo is trying to collect specific examples of driving in snow and ice for the purpose of training their NN to handle snow and ice. So they need specific data on snow and nice. So I think it makes sense to send a small fleet of cars out when roads have snow and nice on them to get the specific data they need. You could collect 1B miles of random data but then you would have to sift through it all to get what you need. Quantity is important but at some point, it is not always better. You can become compute constrained where you have more data than you can handle. Plus, when you collect a lot of random data, most of it will be bad or useless data you don't want. I think targeted data is more useful for ML training.
So yes, I think 10 cars would be sufficient for what Waymo is trying to do.
EVNow
Well-Known Member
LOL. It’s not about how much gig of data you collect but the variety.
All these niche companies have very small amount of data. Basically you need as much data with rain and snow as you do without.
diplomat33
Average guy who loves autonomous vehicles
Keep believing your Tesla myths. Waymo is not a niche company. They don't have a small amount of data. Waymo has a lot of variety, certainly enough for solving autonomy. The fact that Waymo's autonomous driving is years ahead of everybody should prove that point.
EVNow
Well-Known Member
These kinds of proof devoid assertions reminds me of a certain political character …
We are talking about snow etc. Waymo doesn’t have service in any snow areas. How exactly are they “years ahead of everybody” ?
diplomat33
Average guy who loves autonomous vehicles
No, Waymo does not have service in snow areas yet. That's the whole point of the testing in Buffalo, NY. Waymo is testing in snow in order to be able to launch a driverless service in snow areas in the future.
The reason I said Waymo is years ahead of everybody is because Waymo has driverless in the most ODD, more than any other company. Waymo has deployed driverless in day, night, rain, fog, suburb and urban. That's more than any other company has done. Again, the point of the testing in Buffalo, NY is to be able to expand their driverless to include snow in the future.
No, that's not how data science works. Higher sensor resolution than necessary does not overcome lack of data diversity.
We already know Cruise does not have enough data (CEO showed models essentially overfit to SF, still improving in SF when added data from other cities).
So what proof do we have that Waymo has enough data? Do they have a few orders of magnitude more data than Cruise? Have they showed us that their models aren't really improving anymore (especially in prior tested regions like SF) as they collect new data?
diplomat33
Average guy who loves autonomous vehicles
You are confusing two different points I am making. Higher sensor resolution will give Waymo higher quality data. Driving in different cities, in different conditions, combined with simulation augmentation, will give Waymo the diverse data they need.
I think you are trying to change the subject. We were discussing how many cars Waymo would need to collect sufficient data in Buffalo, NY, not whether Waymo has enough data overall. Clearly, Waymo needs more data. That's the whole point of Waymo going to Buffalo, NY! But EVNow asked if 10 cars would be sufficient for Waymo's data collection in Buffalo, NY. I provided my personal reasons for why I think 10 cars would be sufficient for collecting good data in Buffalo, NY, not that Waymo has enough data overall. Of course, we don't know how many cars Waymo will use to collect data in Buffalo, NY. It could be more than 10.
2 reasons off the top of my head.
1) Variety of data: there are at least 13 different types of snow (based off Inuit language translations) and no single location gets them all. Some types pack, some don't, some are sticky, some are slippery, some are more like sand, etc. Temperature, humidity, wind speeds, and altitude all impact what type of snow falls. Snow clearing techniques also change with temperature (salt vs sand vs beet juice).
2) They also need to record how behaviour of other drivers, pedestrians, etc. on the roadway changes for prediction and planning purposes in that type of weather. That includes interacting with large numbers of people during their daily commute who might be snow-blowing driveways, clearing snow/ice from their car, etc. Behaviour, much like laws, are somewhat regional in nature.
Buffalo is on par with Chandler: Mid-sized, wide streets, and slower pace but still populated enough to record a few interesting events.
Last edited:
diplomat33
Average guy who loves autonomous vehicles
Similar threads
- Article
