Hacker News Reader

Comment by greesil | Hacker News Reader

I was biking in SF the other day, pulling my kid in a bike trailer. A waymo vehicle pulled up behind me while I was waiting for the traffic light. I was pleasantly surprised that it waited for me when the light turned green and didn't run me over. Robot anxiety is real though.

Edit: anxiety also because I wasn't sure it would recognize the bike trailer as a real thing. How many bike trailers end up in the training set?

View on HN

Waymo has taken a responsible approach to self-driving (level 5 only, Lidar) over Tesla's collateral reduction approach. Lidar gives you exact distance to obstacles around it. Given those 2 factors, I expect Waymo to have an inbuilt hard-stop for when it is on collision course inside of braking distance.

I expect it to be friendlier to bikers and pedestrians than human drivers.

As a biker, most of close calls are due to impatient drivers who try to overtake me through close calls and when they enter turns without looking first. AFAIK, Self-driving cars do not initiate same-lane overtakes and can be hardcoded to be paragons of patience.

With how hostile American drivers are towards bikers, self-driving cars would need Skynet-esque malice towards us to match those injury rates.

Waymo is level 4, not level 5, and if I remember correctly the CEO said that it won’t be level 5 for decades.

Thats a refreshingly honest statement, especially contrasted with claims by Musk that Tesla Lvl 5 will be easy to do and available to purchase by 2020.

It can't have a total "hard-stop" because otherwise it would break hard for a paper bag or some exhaust smoke.

Does exhaust smoke consistently reflect lasers ? In my experience, a dense lidar will see a subset of the lasers go through such occlusions and get a 2nd return from the truly opaque object behind the smoke.

LIDAR can usually distinguish translucent smoke from solid objects like bikes/pedestrians even before the ML algorithm kicks in.

Paper bags are tricky, because a bag on the ground could just as likely be a solid object that the car should be stopping for. Stopping or slowing down for bags may be a good idea.

> break hard for a paper bag

This is why you have radar as well. The paper bag absorbs minimal energy and you roll through it. (also it's "brake")

For readers that have not watched the TED talk about Google's self driving cars (now Waymo): https://www.ted.com/talks/chris_urmson_how_a_driverless_car_...

It's worth noting that TED talk is from 7 years ago and it's still much better than what a modern Tesla appears to be able to determine.

I would like some kind of visual feedback (maybe countdown timer for how long it has stopped for?) from autonomous car.

Something similar to checking if driver is paying attention and is aware that I'm going to cross in front of their car by looking at direction where driver is looking or a hand gesture.

I really want this, too. I highly depend on eye contact with other drivers for safety.

Interesting! Like a higher fidelity turn signal.

This is a great idea. Disney/Pixar have this figured out, why not turn fiction into reality for a good use case?

With a human I try to make eye contact. At least that way if I get run over I know it's intentional :)

Maybe they were on to something with the Jonny Cab mannequin in Total Recall.

How do you achieve this? I feel like every time I'm crossing the street windshield glare is so severe that I couldn't even be convinced the driver was in their seat.

You could put eyes on the outside of the car.

-try-

This is one of the reasons I wish drivers would stop trying to let me cross the street when they could just drive past. Most of the time I can't see them clearly enough to make eye contact and be sure that they are not just lost/confused (I live in a tourist town with diabolical roads so that's pretty common).

Just look directly into the Lidar. That way we'll know when the uprising is close.

Do not look directly into the lidar with remaining eye

I remember Ford doing some of this work in a research phase a few years ago (by putting real people in hilarious empty seat costumes). I don't recall anything more recent from the major players and it still feels like a missing component to to having vulnerable road users feel comfortable around autonomous cars. https://www.youtube.com/watch?v=KqQyYOPPn7w

OMG that is hilarious! How do we I get one of these costumes?

> I would like some kind of visual feedback

Good idea. Maybe the car can turn on an indicator when it has made the decision it is going to move, something that even gives an idea of where it is intending to move.

you could create some kind of modulated visual signal that activates only on the side that the vehicle intends to move towards?

I know someone at waymo, maybe they can file a feature request. Jonny Cab here we come.

Interestingly, the Waymo vehicles do actually have a screen on top, though I'm not sure what they're planning to use it for. Here's a photo [1].

[1] https://www.wallpaper.com/transport/waymo-division-autonomou...

Yeah same experiences here, already feeling like robot cars are much safer for cyclists and pedestrians than aggressive humans

I’d trust a waymo more than a tesla though

Tesla FSD beta has something like three orders of magnitude more miles driven under full autonomy now, almost all of it by non-corporate users free to post whatever mistakes or mishaps they can find.

Waymo has one (now two) heavily geofenced service areas with a few dozen cars. It won't take highways. It won't enter certain intersections or transit certain neighborhoods.

Just recognize that there's some level of cherry-picking error in this kind of analysis. FSD beta and Waymo both make mistakes (mine curbed a tire once, there's the now-famous video of the car hitting the plastic bollard in San Jose, etc...), but you're going to see far more of those from one system than the other.

> Tesla FSD beta has something like three orders of magnitude more miles driven under full autonomy now

Wait. “Keep your hands on the wheel” drives are NOT fully autonomous. By definition.

Glad to see this is getting thoroughly downvoted.

-> Tesla is actively killing people. -> Tesla has much higher rate of requiring driver intervention, and mostly drives in safer environments.

I really struggle w/ how to communicate to non-industry people how deeply wrong-misleading the Tesla pitch is.

> Waymo has one (now two) heavily geofenced service areas with a few dozen cars. It won't take highways. It won't enter certain intersections or transit certain neighborhoods.

Literally this is why I trust it more, the people running the company are willing to realistically evaluate the limitations of their systems and act responsibly

Totes. "self driving" Teslas are downright scary for cyclists

Tesla self driving is scary for everyone. There's more than just a few crashes and concerns regarding it's capabilities:

https://en.wikipedia.org/wiki/Tesla_Autopilot#Safety_concern...

The Tesla pedestrian AEB systems (and those on any vehicle made in the last 5 years) are pretty good regardless.

They sometimes post videos of situations they deal with, especially around bikers and pedestrians. It's quite impressive how well it handles such cases.

https://www.youtube.com/watch?v=mWvhw1KCmbo

The marketing video vs. the in-the-field video: https://youtu.be/zdKCQKBvH-A?t=833

500k fully driverless miles in Arizona and this is still the only video where it’s actually demonstrated a “failure”. That says something.

In fact, the very same channel (JJ Ricks) has a ton of in-the-field Waymo videos where it works exceedingly well.

> That says something.

Perhaps that not everyone records a taxi ride and posts it on YouTube.

That’s fair. But a lot of them also just try out Waymo for the cool factor and capture the ride (like myself).

Definitely. I still think it is very impressive, I just think they have probably had a few more "bad rides" than this video.

This is a very safe way to fail though, with no vulnerable road users nearby.

But at 16:46 it continues on and moves into traffic before once again failing, this time with people having to go into the oncoming turn lane to go around the Waymo vehicle.

It's notable that most of the failures in that video are human process and communication failures, not failures of the self driving system itself.

It's the kind of thing that could all be fixed in a few days of rehearsals of what to do when cars get into this state. I suspect that the main failure was that rider support had a button to 'pull over', but that the route to pull over was blocked by a cone, so the car failed to pull over, and then rider support seeing the failure probably tried to cancel the pullover (to try again), and it ended up keeping driving.

A simple logic change or even just better training/documentation could have fixed this.

The video contained almost no information. Just a clip of a cyclist cutting in front of a Waymo car and then a cut back to the Waymo employee. It doesn't even show how the car responded.

I think this blog goes into more detail: https://blog.waymo.com/2021/06/beyond-the-bike-lane.html

That still doesn't say anything other than "our data includes interactions with cyclists." As far as I can tell, they don't treat cyclists any different than they do other obstacles on the road.

They do, the model specifically marks cyclist, pedestrian and other obstacles as different, and they discuss how they're modeled differently. Furthermore, they have videos talking about their simulation system, and how they create thousands of "variations" of each scenario, with the "obstacle" doing different things.

That is ridiculous intersection. What is the point of bike lane if it's used as turn lane anyway?

Scania did a neat affordance for this in their (otherwise rather incremental) autonomous AXL concept. A band of LEDs around the vehicle that light up "towards" a pedestrian once the vehicle takes them into account: https://youtu.be/0WN9xvAvEls?t=499

Waymo and cruise vehicles are much safer for bikes than other drivers in my experience. I've been commuting by bike in SF for years, have a near miss every few days but never even close with a waymo vehicle. Every single time I've been hit has been a door, another bike, or a taxi. Taxi drivers in SF seem to see themselves as natural predators for bicycles.

I wonder what the stats are on those bike trailers. I see people using them in traffic and they look plain terrifying: a child dangling out in a flimsy canvas pod behind the cyclists's back, too low for drivers to see when close, surrounded by vehicle wheels... yikes.

There shouldn't be anything inherently dangerous about them, a car shouldn't be getting that close to the back of a bicycle really. I've seen some that have a tall flag-like thing, I guess that's to help show that there's something behind the bicycle

Shouldn't doesn't mean doesn't. Seems to me that common sense suggests they're pretty dangerous to use in traffic, but I'm not claiming that is actually the case. I'm just wondering what the stats are - the only real way to know.

For context, I live in a major city. I've seen people pulling these things round busy multi-lane roundabouts and the like.

Many people seem really careless with their children on the road. Another example is walking out into a crosswalk with a stroller without looking both ways. Yes, it's my responsibility to slow down. But for god's sake don't assume I'm going to!

Those aren't strollers they are traffic testers.

Agreed...

I was thinking about this other day with my 4 month old. We were looking at four or so houses to purchase with a real estate agent. We were having to load her in and out of her car seat in the car repeatedly which is not fun (she cries when you set her in the thing). One of the houses was 1/2 a mile or so away and it crossed my mind that I could just hold her in the back seat while we coast through a few intersections. I then pictured the outcome of a minor accident and decided that it was worth the additional effort to ensure that my kid wouldn't be ejected.

Everything we do is a risk calculation and some people have decided that the risk is worth the convenience or cost.

I suppose my thought was illegal while toting your child around in a bike trailer isn't so there is that factor too...

Those kid trailers seem very dangerous, I ride my bike in cities all the time (NYC and SF) and get very close to reckless drivers. I have to imagine being a passenger in a bike trailer, in a city, is quite a risk

The most common accident from biking is falling. Much of the danger of falls are removed for a child in a proper bike trailer. With a flag and reflectors helping visibility my intuition tells me that it is much safer for children in trailers. Safer than riding their own bike or with a child seat on the adults bike.

Most people I know who regularly bike in SF (for commuting at least), have been hit by a car in some capacity.

I have no idea why parents use those bike with the kids in the back. Why not just use a car? It's not even like those bikes are particularly cheap. The ones I see sell for thousands of dollars.

It must feel very gratifying to be able to say, "hey, you shouldn't have done that" after a driver crushes your child into a paste. Never rely on what drivers "should" or "shouldn't" do.

As a motorcyclist, I'm well versed in not relying on drivers. SMIDSY all day. My point is that it should be safe; if it isn't, something is wrong with the system.

In Germany many bike trailers have a small brightly colored flag on top, so there's an object in your field of view even if it's very close to your car.

There's a subset of cyclists that don't seem to have fear. Not me though. When I take by kiddo of out I'm very careful with route planning. They're harder to see because they're lower down, but the kid literally can't fall out of it if I take a fall. That, and it effectively has a roll cage.

The stats are that probably hardly anything happens with them. Dying on a bike is extraordinarily rare in the first place. For example in my city of some 4 million, 30 people on bikes die a year which seems like such small odds.

Merely because so few people bike in a commuter/traffic heavy setting.

I decided to go ahead and do some ballpark math. about 2% of my city bikes to work:

36/80,000 = 0.00045

80% drive a car and about 250 die a year:

250/3200000 = 0.000078125

So its not quite a full order of magnitude more dangerous, but either way I'd say your odds are pretty low in either case. Even lower if you choose to bike safely, such as taking less busy residential streets vs commercial arteries or even opting for the sidewalk on those busy arteries, wearing helmets and using lights, using defensive bike strategies like avoiding lefts in favor of utilizing a perpendicular street, not being drunk, etc. I bet out of those 36 deaths in my city, none took place on like a 25mph residential street.

I bike around the city often, mostly at night. In the last week I almost got hit by one turning in front of me, and I watched another blow through a red on 9th and Judah. I've also observed them slowing drivers down with erratic behavior. Now I understand why the sudden uptick in volume, but am predicting it gets locked back down soon bc from what I've seen they have much more to learn. I'm very cautious around them. Good for the company and team behind it tho, it's a good fight to wage. Fingers crossed.

Surely you don't solve this problem by recognizing specific objects, you just identify "is there /anything/ in front of me except for road or vehicle travelling at a similar speed to me/within a certain distance of me" and stop if so.

It's been about a year now, but I used to skate around the city all the time -- even then I found the self-driving test cars to be way less treacherous than the human-driven ones!

Pulling your kid in a bike trailer in SF is awesome and admirable, but also a little bit crazy :-)

Yeah it's funny how human psychology works! We know the rational truth is that the AI drivers are millions of times less likely to run us over than human drivers, but we still can't get rid of the image of evil robots going bezerk from Hollywood movies and the (extremely unlikely) possibility of a malfunction in the back of our minds!

I think the problem is that (at least for now), we don't have good mental models of AI drivers and what their failure modes might be.

Dangerous human drivers are likely to be consistently erratic (e.g., drunk), or distracted and oblivious in certain situations (bikes, intersections, lane changes).

For all we know, an AI driver might drive perfectly for three hours, and then accelerate full speed onto the sidewalk.

Uh... people have already been killed by autonomous cars. And they're very new. I don't think it's funny/weird at all that someone would be nervous about being in front of one, without an airbag.

The fact that someone was killed by a car with an Uber computer on board does not mean that people have been killed by autonomous cars. It's easy to draw a bright line between realistic, responsible self-driving tech companies like Cruise and Waymo and the move-fast-kill-people shenanigans of Uber and Tesla. They are really two different categories.

Maybe for someone who is paying close attention, or even works in the industry, this distinction can be made.

And that's valid for now. That a company has not had a major accident so far does not mean that I'm not going to be their first victim.

I love the idea of self driving cars.

But they are also really scary.

They're being brought to us by the same industry that gave us

* BSOD

* "why the heck can't I share my screen right now? can you share your screen? yes? well why can't I share my screen?"

* Your account has been terminated for breaking the terms and conditions. You cannot appeal this decision. We are now deleting 4TB of your life's work.

* An attacker has gained access to your email account. Your life is now ruined.

`But they're safer than human drivers!`

So what. They're new, they're capable of killing you, they're brought to you by tech.

They're f'n scary.

Self-driving advocates will do well to acknowledge this rather than being dismissive.

It's easy to dismiss you because your arguments are nonsense.

Uber's murder car was not autonomous. It was not capable of ever operating without a "safety driver". Therefore it was not an autonomous vehicle. Nor is any Tesla.

Waymo has operated cars with zero humans aboard at various points going back quite a few years. They are legitimate autonomous vehicles.

I'm not concerned about evil robots. A coordinated cyber attack? People intentionally exploiting bugs in perception? Errors in general?

I think the risk of a malfunction is much higher with a robot than a sober human, and probably most legally intoxicated humans.

I don't base it on Hollywood, but in years of programming experience. Anyway, did you hear about the 737 that crashed due to a software bug last month?

> We know the rational truth is that the AI drivers are millions of times less likely to run us over than human drivers

Thank you for a textbook example of begging the question.

> We know the rational truth is that the AI drivers are millions of times less likely to run us over than human drivers

What is that based on? Do you mean the proportions of AI and human drivers?

If you were unpleasantly unsurprised to be run over the other day you wouldn't be posting.

It would be massive news, and we would still hear about it.

(There has been one death so far, with Uber behaving super irresponsibly: https://www.jefftk.com/p/uber-self-driving-crash)

It would likely have it's own HN post rather than just a comment, lol

This is clearly untrue, given the fact that a biker being run over by an at-fault AV would be national news.

Yeah, survival bias.

Reply on news.ycombinator.com