Waymo

[diplomat33]

00:00 Ride start
00:42 Yield to pedestrians at stop sign
01:32 Somewhat blind unprotected right
02:05 Light rail vis and group of cyclists
06:09 Unprotected left
06:52 Safe turn in front of pedestrians
09:58 Tricky-ish unprotected right
10:15 Railroad tracks!!!
14:15 Railroad tracks!!!!!!!
16:05 Slows for trucks
16:30 Misses the turn
16:40 Unprotected left
18:44 Unprotected right at stop
19:38 Misses the turn again
20:48 Unprotected right at stop
21:50 Unprotected left at stop
22:38 Wacky u-turn
23:20 Waymo offroad!
23:35 Not great dropoff/pullover

 

[diplomat33]

0:00 Pull out & introduction
0:19 Asserting right of way on creeper at stop sign
0:42 Passing double parked vehicle
1:03 Yielding to Coco delivery robot
1:34 Nudging around vehicle in crosswalk with pedestrians
3:13 Nudging for parking car
3:48 Unprotected left turn with oncoming traffic (1)
6:29 Unprotected left turn with oncoming traffic (2)
7:37 Uncomfortable speed profile approaching stop sign
9:11 Turn signal at red light for upcoming lane change
9:43 Entering parking lot
10:48 Turn signal at red light for upcoming lane change
15:49 Pull over

 

[diplomat33]

Waymo driverless ride doing an airport terminal drop off in holiday rush hour traffic:

 

[diplomat33]

A busy Sunday evening in downtown Santa Monica in a Waymo. The AV navigates several challenging scenarios with dense pedestrian and vehicle traffic. Taken on November 4, 2023.

0:00 Pull out
2:36 Unprotected left turn with oncoming vehicles
3:12 Destination update during pullover
4:57 Passing indecisively double-parked vehicle
5:51 Right on red
7:14 Unprotected left turn
8:02 Nudging for right turning vehicle
11:56 Right on red with pedestrians crossing
12:41 Right on red with cyclist crossing
13:00 Indecisive lane selection
13:54 Right on red with pedestrian crossing
14:10 Yielding to pedestrians in street
15:26 Nudging for pedestrian near curb & unknown reroute
16:07 Nudging too much for pedestrian near curb
16:59 Right on red with pedestrians near curb
18:34 Right on red with pedestrians crossing
19:34 Unprotected left turn with oncoming vehicles
20:08 Yielding to pedestrians in mid-block crosswalk
21:39 Unprotected left turn
22:46 Slowing for narrow lane
24:04 Unprotected left turn too assertive with oncoming cyclist
26:04 Very long pull over

 

[diplomat33]

And a new JJ Ricks video:

00:00 Interesting turn behavior
01:02 Cross-the-road unprotected left
05:38 Freeway-join fakeout
06:32 Merging with offramp traffic
08:48 Extremely tight three point turn at dead end
10:22 Dog marked as pedestrian on map (pretty cute!)
12:47 Unprotected left on stop, merging via chicken lane
17:20 Entering Downtown Phoenix
18:00 Pedestrian yield right on red
18:40 Awesome yield to scooter
19:42 Stop sign labeled where none exists. Mapping error, or something else?
24:04 Crossing street unprotected, stop sign

 

[spacecoin]

03:15 blew my mind.

Edit: Fantastic presentation as a whole.

 

[diplomat33]

03:15 blew my mind.

Yeah, being able to correctly interpret hand gestures is super important because of situations like that with a construction worker. If the AV does not understand the hand gesture in context, it will likely be indecisive, especially with the hand held stop sign that the AV could read as an order to stop. The fact that the Waymo detected the hand held stop sign but also detected the hand gesture to go and understood that the hand held gesture took priority and it was ok to go, is impressive. It is the difference between the driverless car handling the construction zone fine or stalling and needing remote assistance.

It is a great presentation. Waymo has focused a lot on "total scene understanding" which I think is a key area. I would argue that "total scene understanding" is probably one of the big remaining challenges in autonomous driving. This is because AVs today are actually pretty good at driving when the drivable space and the driving "rules" are normal. But AVs struggle when the drivable space and/or the driving rules suddenly change. There are plenty of situations like construction zones, car accident blocking the road, police blockade, road flooded, city street festival, school crossing etc... where the normal drivable space might be different and the rules are different. The AV needs to take into account the bigger picture and contextual clues to understand what to do.

The other component is what Anguelov calls "openset perception", ie the ability of the perception stack to detect and predict objects outside of its training set. This is key because as the AV scales to more places, it will see new objects outside of its training set. The fact is that you are not going to scale efficiently if humans have to find and train the AV on every new edge case or object before deploying. There are just too many. It will be much more efficient if the AV can simply handle new objects on its own without requiring any human pre-labeling or pre-training.

I believe that once AVs master "total scene understanding" and "openset perception", we will be able to scale AVs wide because AVs will be much more capable at handling those changing scenarios and new edge cases on its own.

 

[EVNow]

Yeah, being able to correctly interpret hand gestures is super important because of situations like that with a construction worker. If the AV does not understand the hand gesture in context, it will likely be indecisive, especially with the hand held stop sign that the AV could read as an order to stop.

Would Waymo understand gestures by another driver ... ? Happens often here at stop signs (people are nicer here) ...

 

[diplomat33]

Would Waymo understand gestures by another driver ... ? Happens often here at stop signs (people are nicer here) ...

Honestly, I can't say for sure. I know Waymo has trained on hand gestures and we've seen examples of construction workers and police officers. I know the lidar has the resolution to detect hand gestures by another driver because we've seen videos of the lidar visualization that show a very high level of detail where you can clearly see the arms and hands of people. Whether that means that Waymo would actually understand gestures by another driver, I don't know. If anyone has an example I would love to see it.

 

[EVNow]

Honestly, I can't say for sure. I know Waymo has trained on hand gestures and we've seen examples of construction workers and police officers. I know the lidar has the resolution to detect hand gestures by another driver because we've seen videos of the lidar visualization that show a very high level of detail where you can clearly see the arms and hands of people. Whether that means that Waymo would actually understand gestures by another driver, I don't know. If anyone has an example I would love to see it.

Apart from understanding the gestures - Waymo has to decide whether to follow the gesture or not ...

speaking of which - how does Waymo decide whether the person making gestures on the road is genuine or a prankster ?

 

[diplomat33]

Apart from understanding the gestures - Waymo has to decide whether to follow the gesture or not ...

speaking of which - how does Waymo decide whether the person making gestures on the road is genuine or a prankster ?

I am just speculating but presumably Waymo takes all the perception info into account to make a decision. That's what "total scene understanding" is. The car takes into account the entire picture, the road, lane geometry, other vehicles, VRUs, stop signs, traffic lights, drivable space, right of way, presence of obstacles, construction zones etc... to make the decision of what to do.

 

[ThomasD]

Wonder how Waymo Winter testing in Buffalo is going?

 

[rbt123]

Wonder how Waymo Winter testing in Buffalo is going?

Hasn't started yet.

They did a bunch of trips in November for a baseline and planned to return in January to capture the snow version.

In addition, it didn't snow in December. Much of the month was +5C.

 

[diplomat33]

A Waymo vehicle that didn't have any people in it was struck by a hit-and-run driver in Phoenix early Tuesday morning, city police said.

The crash happened at the intersection of Central Avenue and Osborn Road around 3 a.m. and the other vehicle involved took off from the scene, according to early reports.

Waymo shared more information about the crash, and said that their autonomous car was going 35 miles per hour and entered the intersection during a green light. The other driver ran a red light at roughly 63 miles per hour and struck the autonomous car before driving away.
There were no passengers or a human driver in the Waymo car, and no injuries where reported in connection with the crash.

Source: PD: Autonomous Waymo car struck by hit-and-run driver in Phoenix

Glad that there were no injuries. I am guessing the human driver was probably drunk. It would make sense since it was 3am and they ran a red light and fled the scene.

Waymo has shown examples of the car stopping at a green and yielding for a red light runner to avoid a collision. Presumably, it was not possible in this case. Or maybe the Waymo did yield and the red light runner clipped the Waymo anyway. We would need to see video to know the details. In any case, based on the info in the article, the human driver would be at-fault and clearly reckless driving (62 mph in a 35 mph area).

 

[Supcom]

Source: PD: Autonomous Waymo car struck by hit-and-run driver in Phoenix

Glad that there were no injuries. I am guessing the human driver was probably drunk. It would make sense since it was 3am and they ran a red light and fled the scene.

Waymo has shown examples of the car stopping at a green and yielding for a red light runner to avoid a collision. Presumably, it was not possible in this case. Or maybe the Waymo did yield and the red light runner clipped the Waymo anyway. We would need to see video to know the details. In any case, based on the info in the article, the human driver would be at-fault and clearly reckless driving (62 mph in a 35 mph area).

Fleeing the scene of an accident with a mobile surveillance vehicle is certainly not a signature characteristic of a sober mind.

 

[DanCar]

Source: PD: Autonomous Waymo car struck by hit-and-run driver in Phoenix

Glad that there were no injuries. I am guessing the human driver was probably drunk. It would make sense since it was 3am and they ran a red light and fled the scene.

Waymo has shown examples of the car stopping at a green and yielding for a red light runner to avoid a collision. Presumably, it was not possible in this case. Or maybe the Waymo did yield and the red light runner clipped the Waymo anyway. We would need to see video to know the details. In any case, based on the info in the article, the human driver would be at-fault and clearly reckless driving (62 mph in a 35 mph area).

60 mph crash. Yes there injuries, but none reported, since the red light runner ran.

 

[AlanSubie4Life]

Curious to see video. This is surprising and seems like a very easy crash to avoid (even for a human).

I wonder if there was a complication here with very limited visibility (blind, reliant on people to comply with the signals) or something. Otherwise it is hard to explain and a Waymo failure.


We’ll see soon hopefully.

Hopefully not a repeat of something like the Cruise left turn failure. Have to avoid the collision! The capabilities in that regard should far exceed a human.

I looked at Streetview (I assume this is the correct intersection) and it looks like if the Waymo was traveling east on Osborn, visibility would be obstructed by the bus stop to an extent. All other directions look trivial to avoid red-light runners at 60mph.

So I assume Waymo was traveling east at 35mph. Still, seems avoidable; headlights should be obvious.

 

[Doggydogworld]

I looked at Streetview (I assume this is the correct intersection) and it looks like if the Waymo was traveling east on Osborn, visibility would be obstructed by the bus stop to an extent. All other directions look trivial to avoid red-light runners at 60mph.

View attachment 1005450

If Waymo was traveling west on Osborn at 35 mph a 63 mpg red light runner coming from the left would only be visible for a second prior to impact. The white building with fancy glass corner in the center of your pic would obstruct the view before then.

That assumes no meaningful braking by either car. If Waymo braked before impact the runner might have been as much as 200 feet away and not visible until Waymo's nose entered the crosswalk.

Fleeing the scene of an accident with a mobile surveillance vehicle is certainly not a signature characteristic of a sober mind.

Depends on the jurisdiction. SF police almost certainly have the tags of the car that hit the woman and flipped her into Cruise's path, yet no arrest has been made. I doubt they even pursued it.

 

[AlanSubie4Life]

If Waymo was traveling west on Osborn at 35 mph a 63 mpg red light runner coming from the left would only be visible for a second prior to impact.

Yes this is possible. Just seems like sight lines are less obscured (to machine vision) in the westward direction (a little more time). As a human I agree that going west could even be worse than going east - it would not be too difficult when traveling east to see through the bus stop especially at night.

If Waymo braked before impact

Yeah that is my concern about what happened. Admittedly I have not tried to go through the math but at 35mph and 65mph for the other car it seems like the car would have to be really close to not be able to avoid a collision simply by proceeding (or stopping short). And if it were really close it seems it would be visible even with the sight lines and the Waymo would just stop short.

Stopping distance from 35mph is at worst 45 feet; that is all the math I have done. I guess that does take 1.75 seconds but not sure that time matters.

This view seems like it would give enough time to react; it gives more than 45 feet to the intersection. And it looks like you can see vehicles 170 feet from your intersection point with them (2 seconds). But yeah it is close. Would have to measure to see how far back you can see. Might be less.

Of course at night one might be able to see rapidly moving headlights via various reflections - though see no obvious useful ones here (and illumination though that would be tough in a well lit city) without seeing the car itself.


Anyway hopefully detailed accident report will be published soon.

I think I agree that with precisely the right timing an accident could be unavoidable if the oncoming car does not slow down, with these sight lines.

Alternative scenario. Seems like the Waymo will be able to transit 75 feet to freedom in 1.46 seconds before the car travels 150ft in 1.57 seconds. Decision making time is zero so is not relevant. Speeding up improves margins.

But certainly depends on the lane the Waymo is in. Fast lane?

In future robotaxis will be programmed to maximize sight lines at all times.

 

[diplomat33]

Really good driverless ride in LA:

0:00 Pick up
0:28 Pull out
4:03 Sketchy unprotected left with oncoming vehicles
6:18 Pull over
6:34 Pull out
6:48 Unprotected right turn with cross traffic
7:39 Slowing for unknown
8:56 Nudge for stopped bus
9:12 Narrow oncoming negotiation
10:56 Awkward trajectory for unprotected left turn
11:40 Visualization zoom out for unknown
14:00 Nudging for pedestrian and open door
14:48 Unprotected right turn with cross traffic
14:56 Nudging for vehicle in parking lane
15:06 Pull over
15:30 Pull out
16:02 Unprotected left turn with cross traffic
22:18 Pull over

The unprotected left at 4:03 was a bit jerky but really assertive in the end. The car saw that the other cars were making right turns so the left most lane was safe to turn into. Well done!

At 9:12, that was a really confident negotiation. There was no awkward "who is going to go" moment. The Waymo was going to move over to let the other car go but when it saw that the other car was yielding, the Waymo just went without hesitation. Nicely done.

Overall, I feel like Waymo is getting almost human-like confidence and assertiveness in certain situations. And it seems better at drop offs too. It actually pulls over to the side.