Waymo's big self-drive car report
arstechnica.com
arstechnica.com
The latest statistics suggest that Waymo's car requires about an order of magnitude fewer interventions from safety drivers. If these systems are a key reason for that or will further boost their safety records, it will be a long while before competitors can catch up.
"Waymo says it has done extensive work to make sure that computer crashes don't lead to car crashes. All of the key systems on its cars—the computer, brakes, steering systems, and batteries—have backups ready to take over if the main system fails."
"Safety-critical aspects of Waymo's vehicles—e.g. steering, braking, controllers—are isolated from outside communication," Waymo writes. "For example, both the safety-critical computing that determines vehicle movements and the onboard 3D maps are shielded from, and inaccessible from, the vehicle's wireless connections and systems."
Does this mean totally isolated with no connection or generally inaccessible but only really at the software level and if you overflow the right buffer with the right bytes you might be able to connect.
However, on this:
"Waymo's car requires about an order of magnitude fewer interventions from safety drivers."
I'm a little skeptical about using interventions per x-distance as a metric for comparison.
There's just too many different things that could skew that. Like driving the same routes over and over. Or more/less stringent rules about when to trigger intervention. Or choosing when not to drive, like poor weather. Or purposely choosing difficult scenarios and edge cases for your routes.
There's not much new here if you've been following what Waymo is doing. Chris Urmson's SXSW talk from 2016 [2] covered most of this material in more detail. This is more of a PR-level version of that content. The new material covers mainly the limitations Waymo is imposing on their system. It works only where they have very detailed maps of the road system, with more information than even StreetView. This is conservative, but not a bad decision for initial release.
They have a button to call customer service. (That's so unlike Google.)
[1] https://storage.googleapis.com/sdc-prod/v1/safety-report/way... [2] hhttps://youtu.be/Uj-rK8V-rik?t=6
I don't think there is anything wrong with linking to a news article like this that is published by a reputable news source such as Ars Technica. Not everyone is going to have the time to read a 43 page report, especially if it's not clear why they should be reading it.
I don't understand why so many news sources skip out on providing references.
That doesn't apply to YC, so when posting here, please give the link to the original source. Sometimes it's necessary to dig through three levels of blog to get to the source. (This seems to be especially true of articles about battery chemistry and surface chemistry ("nanotechnology").)
Obviously they won't be limited to Chandler Arizona forever.
But that's for the future; for now I'm pretty sure every car will have a driver, even if just as a backup/supervisor.
>But it's clear that Waymo is aiming to bring its product to market soon—likely in 2018, if not this year.
Nice weather driving is predicable out ~6 hours and easily 98+% of driving conditions nationwide. If the system simply checks a weather service before driving and says no if there is problems with the route that's fine for version 1.0
Call me cynical but this looks more like Google gave Waymo a deadline and now they’re trying to ship a MVP to keep the funds coming.
This is all anecdotal on both of our parts of course but I wouldn't bet against Waymo on this. They have what looks like a massive head start against the competition here.
I saw one of their minivans a week or two ago. Chandler is mostly "new" construction - New strip malls, newish neighborhoods with nice sidewalks, etc. It's also relatively well-to-do. Average incomes are much higher than in Mesa, which is just to the north.
Phoenix (other side of the Salt River) has nice areas, and it has ghettos. I wonder where the geofence's lines are drawn. Certainly they'd include ASU/Tempe...
I wonder how these would work for passengers who need to get their illegal plant products, dealers of which don't always live in neighborhoods with nice sidewalks. Mostly those passengers didn't tell me what they were actually doing, but sometimes I figured it out.
+ Jealous cap drivers, truck drivers, neighbors can easily put some chewing gum or paint on one or a few sensors. The system will be too brittle to cope with that, so all the time your/the sweet self-driving car doesn't even want to start.
+ Joyriding and pranks. Put your silly little brother or the dog who you've to babysit in the car. Bye bye!
+ Criminality. Just put your drugs on board or a corpse. Or do the companies that own these cars use them as remote cameras all the time?
+ Insurance. Just set up a situation perhaps with multiple cars where the self-driving car slams in your insured car. Use some doll etc.
+ Destruction. Do the same to cause the car falling into the river or crashing into something you dislike.
+ Vandalism. Graffiti etc. will be recognized much later than with a human driver.
+ Theft of sensors or other car parts can be orchestrated on a location the car first drives to.
In Waymo's case, they can solve or alleviate these problems with live cameras and strict passenger screening process.
I'm not arguing that it is insurmountable. I'm just saying that it's too early. Most of these things have not been worked out, which seems unnecessary.
Tesla's "tunnels": I assume you are referring to elon's Boring Company? How many decades of tunneling do you think that will require before it even covers a fraction of routes people want to take?
Waymo has real world usage, both from self-driving cars which already use real streets, and from their worldwide street mapping fleet.
Besides, real world data is only as good as you can capture it, and Teslas are much more limited in that regard compared to Waymo's LIDAR.
If all our movements will be subject to continuous monitoring and approval, then SDC's won't be anything better than taxis or planes.
disclaimer: google employee, but no knowledge of waymo things.
Maybe their "onsight" capability is just very slow and the first Waymo to meet a change has to onsight it at hardly more than roomba speed (or worse: mturk?), sharing the knowledge ("beta") with the Waymos that come later, but then you still have to reliably detect that your model is outdated while driving regular speed while all false positives would considerably harm the relationship with other road users.
This reminds me darpa grand challanhe and urban challenge difference. The cars that were built for the desert wouldnt work great in an urban environment.
Over time these restrictions will surely go away.