Waymo issues software and mapping recall after robotaxi crashes into a pole
theverge.comIf self-driving vehicles are tested on public streets, all data generated should be public. Companies should be cooperating, not competing, on "don't hit telephone poles" and "don't run over children".
Sounds cool in theory, but much less than useful in practice. The hardware and file formats are different enough that the data would be mostly useless to us (I work at a self driving-ish company).
The models are predicated a specific set of hardware ( which camera, what fidelity LiDAR, but even something as simple as “what’s the frame rate” can have a difference)
That might not be too big of an issue. There are papers that enable reusing data of different modality by filling blanks.
Should all Google Street view data be publicly available?
After some sleuthing and Geoguessring, here's the alley where it happened:
https://maps.app.goo.gl/UNUzQn686ESYWbfo7
Further south of the pin, by the section of the alley with stripes on the pavement behind the garages for 842 N 6th Ave (but not on 6th or 7th, on an unnamed alley between the two).
Let me repeat - it's an alley. The Google Maps car didn't even go down that road (though it looks to have been under construction when the Maps car went through).
Without exclusing Waymo (they had their car do something dangerous and stupid) this is the kind of pseudo-off-road parking lot/driveway/construction zone nav stuff that's really hard to get right, and almost requires AGI.
I think the real error was not the damage score but the planning algorithm that directed it to drive down and to continue through that alley.
I think we'll soon get to (if we're not there already) a form of level 2 driver aids or level 3 geofenced self driving (highway only?) that's safer than average human drivers. I think we're a long way from self-driving cars that will assign a low damage score and drive over an empty cardboard box in an unmarked, unmapped private alley, and we may never get there. But that doesn't mean Waymo can't or shouldn't exist, it means they need to shut down the car and delegate to a human when they're stuck and not not on public, mapped, confirmed clear roads. Maybe that means it can't pick you up from the back of the Chic Fil A parking lot or the entrance to the mall that's an island in a quarter mile of private parking lots and you have to go to the nearest parking spot on the actual road, but if the alternative is assigning damage scores to stuff in alleys that's probably for the best.
I'm sure they did it out of caution, considering their excellent track record. Love this company, would like to see them expanding.
I'm interested about the legal requirements of Waymo filing a recall notice with the NHTSA. That is, all the affected cars were actually owned by Waymo - could they have just updated the cars anyway, by themselves.
To emphasize, I'm genuinely curious. I don't understand how the recall notice process works if your product isn't owned by anyone else but you.
Product recall insurance is a relatively common rider on general liability insurance. I don’t know if that’s what drives Waymo’s decision here but it’s pretty common for such policies to require the insured to play nice with regulators and not pull any stunts, otherwise they’ll fight any payout.
I don't think the economics will work (for quite some time). They have ~670 cars in total, and still have seriously negative cash flow.
This is the most interesting part, in my opinion:
> Waymo’s recall was deployed by the company’s engineers at the central depot where the vehicles return for regular maintenance and testing. It was not through an over-the-air software update, like some of Tesla’s recent recalls.
I'd be interested to learn more about why the updates are manual, and also whether the map data is fully local to the vehicle. Tesla obviously does the polar opposite of this, and it seems to have at least some degree of success, but Tesla's approach has always seemed like it would be subject to some bad potential failure modes in my mind.
How much data does this amount to? Gigs? Terabytes?
On the same note, I'm curious about what data gets pulled from the map versus sensor data. The car seems to have used map data instead of sensor data (unless I'm misunderstanding?). Whether there's a curb seems to be exactly the sort of thing you could rely on sensors for, mostly because you also already need to look for obstructions which necessarily can't be in map data.
It's a shame there are so many trade secrets involved, because it would be especially fascinating to get more details on their entire system everything from advanced topics like how they blend ML with traditional rule-based logic to how they manage data like maps (versioning, distributed updates, etc.)
Agreed. They publish some impressive research on https://waymo.com/research/, but are very tight-lipped about many things. I've always wanted to know how they maintain maps and what their in-vehicle compute hardware looks like. They don't want to share anything beyond high level blog posts.
Maybe it's about validation, not necessarily update data. They do seem to take safety more serious than Tesla does.
It sounds weird to say, but updates being applied only in-garage doesn't mean they're not going over the air. However, updates can include software for between dozens and hundreds of individual processors. That makes them quite large. The real data pigs are driving logs and map data though, which are usually serious storage constraints in autonomous vehicles.
Yeah, I can definitely understand that it's probably far faster (and cheaper, and reliable) to transfer the data in the shop with a USB cable in both directions.
In fact, it's probably fastest to have a small bank of hard drives that you can physically pull from the vehicle and swap out for a fresh one. You could probably pack a few hundred TB of storage into a unit that could fit in a briefcase and have it loaded with fresh map data and lots of room for logs.
You've possibly misinterpreted what I said. To gesture vaguely some more, it's not necessarily about what's fastest, but rather what involves the least human intervention. Wish I could be more clear.
Could be a lidar failure, or potential failure, that needed a firmware update to address.
"The update corrects an error in the software that “assigned a low damage score” to the telephone pole."
What did it get classified as? What's a Waymo allowed to hit?
My best guess is they're lying, or they model curbs to have a "low damage score" so its okay if the ai thinks theres a 10% chance a path will hit a curb.
I'm excited for self driving cars, but I have reservations about a system that has to hard code "don't hit a telephone pole". It reminds me of this skit https://www.youtube.com/watch?v=3m5qxZm_JqM
> My best guess is they're lying
It would be extremely silly to lie about an implementation detail that doesn't matter in an official recall filing with an agency currently investigating you for your reporting of incidents like this.
It could also just be everything gets a damage score, including reflection artifacts that aren't really there or paper garbage in the road, and in this case the pole was identified incorrectly, given a low score, and the car thought it could keep driving.
The verge seems to quote the filing, it would be nice if they had linked it. I can't find it from searching.
Cruise withheld details (lied by omission) about the person they ran over...
Cars hit "objects" all the time without issue. Bits of paper flying through the air, bags, cups, steam clouds, and other small objects are very common sources of "motion planning collision" in urban environments.
In lidar it could have looked like a tall traffic cone or a one of those skinny flexible posts that line the boundaries of some bike lanes. Those things cause much less damage. Telephone poles are obviously much thicker but in lidar space the two can look more similar especially if the cloud is very sparse or the sensor is broken.
I kind of love the idea that Waymo sees the world as an RPG, and that mundane objects have damage statistics. I hope Tesla knows I do poision damage.
Other road users could be classified on the lawful/chaotic and good/evil axes.
Well-behaved vehicles are lawful good. Skateboarders are chaotic good. Drunk drivers are chaotic neutral. Road ragers are chaotic evil. Chaotic users are less predictable, and evil users require active evasion.
The robocar "trolley problem" appears to be real.
I assume Waymo is continuously improving the safety of their software. It's bizarre to classify two of these updates as "recalls" but not the others.
"Recall" really needs to be retired as a term.
It’s a term of art; as such, it’s probably here to stay.
But, in this case, it's not a term of art, it's a specific legal definition. It's a "recall" because Waymo filed a recall notice with the NHTSA.
term of art - noun noun: term of art; plural noun: terms of art
a word or phrase that has a precise, specialized meaning within a particular field or profession. "‘public domain’ is a term of art in copyright law and you shouldn't throw it around if you don't know what it means"
What'd be the proper term then? In french when, for example, many cars are sent back to the dealerships (or factory) to fix something, it's called a "campagne de rappel ("recall campaign").
P.S: I've got nothing against "recall" being recalled so that it can be fixed ; )
Recall is unfortunately interpreted as the fix, “physically send back to factory”, where as it is actually a term of art meaning a public notice of a safety-endangering problem that the manufacturer is responsible for fixing, replacing, or refunding for free.
While I have previously agreed that it should not be necessary to change the term, it is unfortunately confusing to laypeople and intentionally misconstrued by bad actors to downplay safety faults. I propose it should be replaced with: “Safety Defect Notice” which is a fairly precise description of the underlying concept and can not be misconstrued to be about the “fix”.
I think that's the problem. "Recall" to most people means "I need to take my car to the dealership, and they will fix something dangerous with the car and give it back to me".
Whereas, may recalls these days are just "Your car performed an over the air update while you were sleeping".
Maybe we need the concept of "physical recall" and "software recall".