Waymo

[MP3Mike]

Waymo blog on how their remote assistance works with some examples of the remote assistance tools that I don't think we've seen before:

Fleet response: Lending a helpful hand to Waymo’s autonomously driven vehicles

The Waymo Driver autonomously navigates tens of thousands of rider-only miles across San Francisco, Phoenix, Los Angeles, and Austin daily. It can navigate common scenarios, like adhering to a crossing guard directing traffic, as well as more unique interactions like avoiding a swerving vehicle...

waymo.com

I thought it was pretty informative.

I think this confirms that they don't count it as a disengagement even if "Fleet Response" directly tells the Waymo where to go:

Fleet response can influence the Waymo Driver's path, whether indirectly through indicating lane closures, explicitly requesting the AV use a particular lane, or, in the most complex scenarios, explicitly proposing a path for the vehicle to consider. The Waymo Driver evaluates the input from fleet response and independently remains in control of driving.

It is like throwing your hands up and asking your passenger how you should proceed. And after they give you instructions you just execute them.

 

[flutas]

Waymo blog on how their remote assistance works with some examples of the remote assistance tools that I don't think we've seen before:

Fleet response: Lending a helpful hand to Waymo’s autonomously driven vehicles

The Waymo Driver autonomously navigates tens of thousands of rider-only miles across San Francisco, Phoenix, Los Angeles, and Austin daily. It can navigate common scenarios, like adhering to a crossing guard directing traffic, as well as more unique interactions like avoiding a swerving vehicle...

waymo.com

I thought it was pretty informative.

I'm honestly shocked that their lead image has errors in the rendered scene when compared to the image of it.

Makes me wonder if it's their lidar data feed or if it's something else that's delayed.

These are in no particular order:

• Jeep in front of the Waymo is at a different angle.
• Jeep in front of the Waymo is occluding the car in the scene, not in the camera.
• Car visible to the right of the Jeep in the scene, not in the camera.
• Scooter rider is behind the police on the left in scene, camera he's in front.
• Cop on the left's arms are a different angle.
• Pedestrian is in the crosswalk in camera, edge of the road on scene. (legs are different position as well)

Based on the cop on the left's arms. If he's moving them around like a normal traffic control cop would it seems like maybe a half-second delay?

 

[diplomat33]

I think this confirms that they don't count it as a disengagement even if "Fleet Response" directly tells the Waymo where to go:

No, Waymo does not count fleet response as a disengagement since the autonomous driving was never disengaged. The Waymo Driver was still driving the car the whole time so it is not a disengagement. But IMO, fleet response should count as an intervention but not a disengagement since a human did help the car but did not disengage the system. But if we do count fleet response as interventions, I think we should only count actual fleet responses, not every time Waymo calls fleet response, since there are cases where Waymo calls fleet response but then resolves the issue on its own without fleet response doing anything. IMO, that is not an intervention since a human did not actually do anything to resolve the situation.

 

[diplomat33]

Based on the cop on the left's arms. If he's moving them around like a normal traffic control cop would it seems like maybe a half-second delay?

Yeah, there is likely a delay caused by needing to send the data over the network to fleet response. That is why fleet response can only be for non safety critical issues. The Waymo Driver needs to always handle safety critical issues on its own. It would be too risky to rely on fleet response for safety critical issues with a half second delay. That's basically when you know it is ok to remove the safety driver. When the autonomous driving can handle safety critical issues on its own at a rate safer than humans then you can remove the safety driver and rely on fleet response for the non safety interventions.

 

[flutas]

Yeah, there is likely a delay caused by needing to send the data over the network to fleet response. That is why fleet response can only be for non safety critical issues. The Waymo Driver needs to always handle safety critical issues on its own. It would be too risky to rely on fleet response for safety critical issues with a half second delay. That's basically when you know it is ok to remove the safety driver. When the autonomous driving can handle safety critical issues on its own at a rate safer than humans then you can remove the safety driver and rely on fleet response for the non safety interventions.

That delay would cover both scene and video though. Plus they are demoing it in a frozen state, wouldn't that make the most sense to use the most-accurate timestamp of the data?

To me this says their scene recognition isn't running as fast as we all thought it would be, after all their lidar is only running at 20fps(hz) and will likely limit it to about that.

 

[stopcrazypp]

Yeah, there is likely a delay caused by needing to send the data over the network to fleet response. That is why fleet response can only be for non safety critical issues. The Waymo Driver needs to always handle safety critical issues on its own. It would be too risky to rely on fleet response for safety critical issues with a half second delay. That's basically when you know it is ok to remove the safety driver. When the autonomous driving can handle safety critical issues on its own at a rate safer than humans then you can remove the safety driver and rely on fleet response for the non safety interventions.

That delay would cover both scene and video though. Plus they are demoing it in a frozen state, wouldn't that make the most sense to use the most-accurate timestamp of the data?

To me this says their scene recognition isn't running as fast as we all thought it would be, after all their lidar is only running at 20fps(hz) and will likely limit it to about that.

Yeah the delay isn't the issue, since after all, this is supposed to be a "snapshot" in time, not a realtime feed. The bigger problem is the video feed isn't synced with the data feed in the "snapshot". That can be a big problem as it may cause a misunderstanding of the scene.

I do wonder if it's a video/photo editing glitch for this particular PR pice (in which case it's less serious) or if this is really the actual UI the operators actually see (Waymo makes it sound like the latter).

 

[Dewg]

It is like throwing your hands up and asking your passenger how you should proceed. And after they give you instructions you just execute them.

You should watch The Amazing Race. The number of times the taxi driver doesn't know where to go and the contestants are yelling at them with instructions...

 

[Daniel in SD]

I'm honestly shocked that their lead image has errors in the rendered scene when compared to the image of it.

Makes me wonder if it's their lidar data feed or if it's something else that's delayed.

These are in no particular order:

• Jeep in front of the Waymo is at a different angle.
• Jeep in front of the Waymo is occluding the car in the scene, not in the camera.
• Car visible to the right of the Jeep in the scene, not in the camera.
• Scooter rider is behind the police on the left in scene, camera he's in front.
• Cop on the left's arms are a different angle.
• Pedestrian is in the crosswalk in camera, edge of the road on scene. (legs are different position as well)

Based on the cop on the left's arms. If he's moving them around like a normal traffic control cop would it seems like maybe a half-second delay?

Watching the video of the actual remote interface I don't really see any significant lag. This could just be an artifact of a fun demo mode (obviously you can't actually pause real life).

 

[enemji]

Much like phone-a-friend, when the Waymo vehicle encounters a particular situation on the road, the autonomous driver can reach out to a human fleet response agent for additional information to contextualize its environment. The Waymo Driver does not rely solely on the inputs it receives from the fleet response agent and it is in control of the vehicle at all times. As the Waymo Driver waits for input from fleet response, and even after receiving it, the Waymo Driver continues using available information to inform its decisions.

Who is the Waymo Driver? Oh yeah. It’s the sofware. Why do they use normal English words and make a mess of comprehension?

Waymos are designed to operate without the need for human input, except the times where it needs human input.

 

[eli_]

Impressive! And they sent a second car in after it so that one can get itself stuck.

 

[JulienW]

Impressive! And they sent a second car in after it so that one can get itself stuck.

My God a finder bender happened in the US yesterday. Thank goodness no other wrecks were reported or believed to have happened in the US. I heard two teenagers almost bump into each other in a mall parking lot but thanks to an alerting bystander that TRAGEDY was averted. Stop the presses and send everyone available to the horrific scene since no human has EVER hit and mauled a pole. This is likely the very first pole hit in history. We need to put an end to this before another pole is splintered. My confidence in tech is shaken and fear and empathy is extended to all the polls out there. Thank you for your service and bravery to all the poles in this mad world of RTs. Remember only human drivers can protect polls and RTs are out to disable our world one splinter at a time. Just think in another 10 to 20 hits that pole may fall in the line of duty and there will be one less pole on earth THANKS to RTs.

Send your contributions to "Save the Poles and Stop the Robotaxi rampaging" fund.

 

[MP3Mike]

My God a finder bender happened in the US yesterday.

But Waymos have multiple LiDAR units, as well as radars and cameras, so that they should never hit a stationary object. So, the news is that even with all of those sensors Waymo still can't avoid hitting a power pole...

 

[enemji]

Impressive! And they sent a second car in after it so that one can get itself stuck.

Waymo was assigned The Trolley Problem.

 

[kabin]

Waymo blog on how their remote assistance works with some examples of the remote assistance tools that I don't think we've seen before:

Fleet response: Lending a helpful hand to Waymo’s autonomously driven vehicles

The Waymo Driver autonomously navigates tens of thousands of rider-only miles across San Francisco, Phoenix, Los Angeles, and Austin daily. It can navigate common scenarios, like adhering to a crossing guard directing traffic, as well as more unique interactions like avoiding a swerving vehicle...

waymo.com

I thought it was pretty informative.

In the article's header video, the most impressive item to me was how quick Waymo processed and evaded the head-on pickup crossing into path. V12 FSD might have initially responded after the pickup either struck or passed the ego.

 

[kabin]

I'm honestly shocked that their lead image has errors in the rendered scene when compared to the image of it.

Makes me wonder if it's their lidar data feed or if it's something else that's delayed.

These are in no particular order:

• Jeep in front of the Waymo is at a different angle.
• Jeep in front of the Waymo is occluding the car in the scene, not in the camera.
• Car visible to the right of the Jeep in the scene, not in the camera.
• Scooter rider is behind the police on the left in scene, camera he's in front.
• Cop on the left's arms are a different angle.
• Pedestrian is in the crosswalk in camera, edge of the road on scene. (legs are different position as well)

Based on the cop on the left's arms. If he's moving them around like a normal traffic control cop would it seems like maybe a half-second delay?

It's a grainy wide angle video to be drawing many conclusions from but the Jeep looks okay to me given UI update delays and vehicle movements.

Waymo likely still has enough info to know a moving object is behind the jeep. That's the purpose of redundancy. All systems aren't required to reliably process the scene.

 

[flutas]

It's a grainy wide angle video to be drawing many conclusions from but the Jeep looks okay to me given UI update delays and vehicle movements.

Waymo likely still has enough info to know a moving object is behind the jeep. That's the purpose of redundancy. All systems aren't required to reliably process the scene.

I mean, a Waymo literally crashed into a telephone pole yesterday. Somehow that redundancy didn't save them there.

I also love how people instantly jump to defend Waymo when I just point out something. I was literally just saying that there's a delay with their data feeds (I've seen the same discussion with FSD and it's reaction time). Somehow that turns into "Waymo is clearly going to murder everyone on the planet trolley problem style and it must be stopped" and everyone needs to attack that position.

 

[ItsNotAboutTheMoney]

I mean, a Waymo literally crashed into a telephone pole yesterday. Somehow that redundancy didn't save them there.

Becoming more human-like every day.
(I have been negatively impacted (metaphorically) more than once due to such collisions).

 

[Doggydogworld]

I think this confirms that they don't count it as a disengagement even if "Fleet Response" directly tells the Waymo where to go:

Definitely should not count as a disengagement. Or an intervention. Request for confirmation/assistance is a completely different metric.

It is like throwing your hands up and asking your passenger how you should proceed. And after they give you instructions you just execute them.

Never execute the passengers!!!
(Well, maybe the really annoying ones....)

Waymo should not have hit the pole, full stop (ha). There is no excuse. That's why you use lidar, to detect stationary things and avoid them. This is an unforgiveable failure of core safety systems.

Speaking more generally, Waymo has regressed in recent months. It's not just "more miles = more incidents". They now make mistakes they never made before. I've previously guessed they're giving ML too much control. But do they really now let ML override basic "don't hit this big object" guard rails? Hard to believe, and yet..... If they really want to build/restore confidence they need to publicly explain this very basic screwup. Unfortunately I expect them to sweep it under the rug, as usual.

 

[stopcrazypp]

But Waymos have multiple LiDAR units, as well as radars and cameras, so that they should never hit a stationary object. So, the news is that even with all of those sensors Waymo still can't avoid hitting a power pole...

And this time unlike the tow truck incident, there is no trajectory prediction involved. It also doesn't look like it was a slight misjudgement of distance (like the curbing that happens with FSD during turns), it just straight up smacked right into it (enough to bend the bumper and front grille in, not just a slight parking "love tap"). From the video, the visibility seemed excellent, so occlusion of the pole by other objects or fog doesn't seem to play a role. Would be interesting to see why it happened.

 

[Doggydogworld]

The clip at the top of the blog entry shows Waymo obeying hand gestures to turn right with no help from Fleet Response. But watch the visualization of the step van behind it -- they show it cutting the corner onto the sidewalk then rolling over on its side!