Autonomous Car Progress

[Jeff N]

The only "benefit" Waymo and Cruise have is lots of investment in remote *human* monitoring capabilities. They've built lots of software and hardware to have remote human decision makers. Any deviation from the intended path (while on laned roads, not in parking lots) will likely require remote human approval. Waymo and Cruise cars have no real "intelligence" about passing double parked cars (among so many other situations). They simply don't have the breadth of data to deal with these cases, as Tesla does.

I doubt that you are really making this claim seriously but, just for the record, I’ve taken late night rides in the otherwise empty Cruise EVs in San Francisco and have watched them quickly and seamlessly handle multiple double-parked cars.

One time involved a double-parked car waiting to take the place of another car that was parallel parked just ahead of it and was about to leave (although this was not obvious to me when my Cruise car pulled up to the scene). After a second or two, the Cruise car quickly began steering around the car in the road and then stopped when it sensed that the parallel parked car was about to drive away — which it did. The double-parked car then moved in parallel to the curb to park and the Cruise car simultaneously moved back into the normal lane slot and continued on its way. All of that maneuvering was precise and happened quicker than any human remote controller would have performed. Other, simpler double-parking situations, were immediately recognized and steered-around gracefully without stopping.

 

[Bladerskb]

I find it funny that people are still discussing the Cruise and Waymo approaches after seeing Tesla's progress to 10.12

The Cruise and Waymo business models are a clear dead end. Big, diverse, and real data are a prerequisite.

what's the update on your multiple 6 months till robotaxi predictions that have consistently failed since around October 2020?

 

[ThomasD]

So how does it know that a line of cars is stopped for a road closure due to an accident and not to proceed around the stopped cars vs double parked cars that it can go around

 

[diplomat33]

So how does it know that a line of cars is stopped for a road closure due to an accident and not to proceed around the stopped cars vs double parked cars that it can go around

How do humans handle that situation? You train the software to do the same thing. You train the perception stack to detect cues like a road closure sign, cops waving at cars to go around, other cars going around. You train the prediction to predict what other vehicles will do. And then you train the planning stack to figure what path to take based on what the perception and prediction stacks say.

 

[Bladerskb]

Blah blah blah, another year, same old story about Waymo Gen XYZ:

I watch Waymo ride videos from time to time, and they're still driving through the most boring and easy to drive areas (with or without safety drivers), avoiding anything approaching the easiest fsd beta drives.

15 years of development and the best sensors still trip up at any signs of cones.

Its funny that you say that because in these "easy drives" Tesla FSD still falls way short of Waymo and its not even close. for example in Phoenix, not only does Tesla require a safety disengagement from time to time in comparison videos, it actually drives dangerously consistently. Waymo on the other hand is perfect every time.

For example in the latest waymo vs Tesla FSD comparison. On the same exact street, making similar right turn. Tesla turns into the opposing lane, scaring both drivers.
Additionally through out the ride, tesla drives in the middle of the road, sometimes driving 25% in the opposing lane. Stopping on the wrong place, being confused at an intersection, not knowing what to do, missing speed bumps and not reacting until the last second. I could go on and on about this one drive. Yet the Waymo didn't exhibit none of these issues.

So even on your so called easy drives. FSD beta falls way short. Even on a one drive sample, let alone a 1,000 or 10,000 drive sample.

If you take a look at SF, if you look at the latest video from AIDRIVER. It took just 10 miles for there to be a safety disengagement at the very end of the video. 10 miles! In the same video you see 2+ Waymos driving. The same city that Waymo has gone driverless in and will soon be public in the coming weeks.

 

[powertoold]

I doubt that you are really making this claim seriously but, just for the record, I’ve taken late night rides in the otherwise empty Cruise EVs in San Francisco and have watched them quickly and seamlessly handle multiple double-parked cars.

One time involved a double-parked car waiting to take the place of another car that was parallel parked just ahead of it and was about to leave (although this was not obvious to me when my Cruise car pulled up to the scene). After a second or two, the Cruise car quickly began steering around the car in the road and then stopped when it sensed that the parallel parked car was about to drive away — which it did. The double-parked car then moved in parallel to the curb to park and the Cruise car simultaneously moved back into the normal lane slot and continued on its way. All of that maneuvering was precise and happened quicker than any human remote controller would have performed. Other, simpler double-parking situations, were immediately recognized and steered-around gracefully without stopping.

Our assumption is that there was no human involved in making the decision whether or not to pass the double parked cars. How do we know for sure? After all, the Cruise car can ping the remote driver well in advance, making it seamless for the passenger. It's not a simple decision in many cases.

I'd only believe it once we see more examples of it happening. Does Waymo and Cruise even claim that all the decisions to pass double parked cars are made by their planner?

 

[powertoold]

I found this. I'm done discussing AVs again, lol. Tesla is now the clear winner in this space. I'll come back in a few months once Tesla releases V11.

But there's also a whole vocabulary to be crafted to explain the unfamiliar space-time of teleoperation. Today Waymo's "fleet response" team members monitor vehicles in the field from a central command center in Chandler, Arizona, and "weigh in with an extra set of eyes" whenever an AV meets a surprise on the road. These operators don't steer but approve AV-planned detours like driving around a double-parked vehicle blocking a one-lane street. The trucking industry is also experimenting with remote human control. Starsky Robotics, a startup, is developing big rigs that drive themselves on the highway but when traveling on local streets are steered by human operators in a distant control center. What do we call this scheme, which resembles the way harbor pilots take control of big ships as they approach the coast? Similarly, but at a much smaller scale, MIT's Sangbae Kim builds robots that move over terrain under Al-guided locomotion, while remote pairs of human hands take control of arms and cranes during fine-grained manipulations of the world in front of them. None of the words we use now-self-driving, driverless, or autonomous captures what's important or distinctive about these human-machine. hybrids. But they'll need good labels if we are to understand what to expect of them out there.

Originally published: August 11, 2020
Author: Anthony M. Townsend

 

[Daniel in SD]

I'll come back in a few months once Tesla releases V11.

Any quantitative predictions?

 

[Terminator857]

Any quantitative predictions?

Prediction after V11 releases will be "I'll come back in a few months" for another future release. Same story since 2016? After HW4 is on the horizon: I'll come back after HW4. After HW5 is on the horizon: I'll come back after HW5. I use to think Tesla was ahead in the 5 year time frame. Now the race with Waymo looks like a tie to me. If Tesla pulls out some impressive releases then likely Tesla will pull out ahead in this long race.

 

[diplomat33]

Now the race with Waymo looks like a tie to me.

What race are we talking about?

If you mean the race to deploy actual robotaxis, then Waymo would be ahead since they have deployed public robotaxis while Tesla has not deployed any public robotaxis yet. And Waymo will soon open up their driverless rides to the public in SF and downtown Phoenix making their lead even bigger, if we are counting number of robotaxis.

If you mean the race to "solve FSD", I think Waymo would still be ahead. Sure, if you only look at a few videos where both Tesla and Waymo drive around Chandler with no interventions, then it might look like a tie. But that is a very superficial analysis. If you look deeper, I think Waymo is ahead in FSD capabilities. Just the fact that Tesla still has a driver while Waymo has been able to remove the driver, says a lot about the reliability of Waymo. Even on "zero intervention" drives, FSD Beta makes mistakes that Waymo does not make like getting into the wrong lane, taking turns in a jerky way, being very hesitant or almost hitting static objects. Waymo rides are much smoother. Waymo can handle emergency vehicles. I don't think Tesla has programmed that problem yet. Waymo can pick up a person and drop them off with no driver. FSD beta will often quit before the destination, leaving the driver to finish the trip. Waymo has also done a lot of work with Prediction, especially with social interactions with pedestrians and cyclists. For example, Waymo has done a lot with "key points" to predict body language of pedestrians, cyclists, skateboarders to anticipate behavior. Those are some areas where I feel Tesla is behind on the FSD part. Waymo has also mentioned that the 5th Gen can accurately track "hundreds of objects" at say a busy intersection. I know FSD beta can track a lot but I doubt it can accurately track that many objects.

Some references:

How we ensure the Waymo Driver operates safely in San Francisco

Last year, Waymo expanded its testing in San Francisco and began welcoming new riders in the city. As we continue to expand our operations providing rides to more San Franciscans, we’re sharing even more on how we evaluate the safety of our technology and operations in the City by the Bay.

blog.waymo.com

Utilizing key point and pose estimation for the task of autonomous driving

Historically, computer vision relies on rigid bounding boxes to locate and classify objects within a scene; however, one of the limiting factors in detection, tracking, and action recognition of vulnerable road users, such as pedestrians and cyclists, is the lack of precise human pose...

blog.waymo.com

 

[diplomat33]

Mobileye just shared this little clip of their vision-only AV navigating the streets of Tokyo to show off their ability to quickly map and deploy in new areas:

https://twitter.com/x/status/1534516921030295552

 

[diplomat33]

Waymo and Cruise cars have no real "intelligence" about passing double parked cars (among so many other situations).

Here is proof you are wrong:

https://twitter.com/x/status/1479487991450120196

Waymo safely and smoothly maneuvers around the double parked truck in a tight spot. It is all done by AI, no remote assistance, no human intervention. And note that the Waymo perception understands that the cone means the truck is double parked for deliveries. So Waymo has AI to interpret cues like cones to know when a truck is double parked or not. Also note the last sentence in the video: "we are deploying neural networks that learn from human preferences". So your claim that Waymo has no real intelligence for cases like double parked vehicles is wrong. They have intelligence that understands cues like cones and can handle these cases.

Frankly, I am not convinced FSD Beta would be able to handle that scenario that well. For one, Tesla lacks peripheral sensors so it probably would need to awkwardly creep around the truck to see if the path is clear. It is also a tight spot. FSD beta would probably hesitate. Second, FSD beta lacks the AI to understand that the truck is parked so it would probably just sit there waiting for the driver to tap the accelerator pedal to go.

 

[EVNow]

House committee sounds alarm on rising U.S. traffic deaths

A U.S. House of Representatives subcommittee on Wednesday held a hearing on soaring American traffic deaths and efforts to build safer roads, calling it an urgent crisis.

www.reuters.com

 

[diplomat33]

House committee sounds alarm on rising U.S. traffic deaths

A U.S. House of Representatives subcommittee on Wednesday held a hearing on soaring American traffic deaths and efforts to build safer roads, calling it an urgent crisis.

www.reuters.com

Not to get political but considering the hyper partisan climate in Congress, I doubt anything will happen. But I personally would love to see good federal regulations that define AV safety and promote AV deployment. I think it would really help speed up the deployment of safe AVs which in turn would help reduce traffic deaths.

 

[Goose66]

So how does it know that a line of cars is stopped for a road closure due to an accident and not to proceed around the stopped cars vs double parked cars that it can go around

If you're talking about Tesla, in my case the answer is "it doesn't." It's constantly trying to go around "parked" cars that aren't parked at all. They may be waiting for oncoming traffic before turning left or they may be waiting in a long line of cars for a light to change and the Tesla just can't see around the one in front of it. It will even cross over a double-yellow line into oncoming traffic to around a "parked" car, or even a school bus!

 

[Jeff N]

Frankly, I am not convinced FSD Beta would be able to handle that scenario that well. For one, Tesla lacks peripheral sensors so it probably would need to awkwardly creep around the truck to see if the path is clear. It is also a tight spot. FSD beta would probably hesitate. Second, FSD beta lacks the AI to understand that the truck is parked so it would probably just sit there waiting for the driver to tap the accelerator pedal to go.

FSD beta certainly tries to decide if a vehicle ahead is parked. I’ve had it go around a double-parked car a number of times without waiting for me to tap the accelerator.

Here’s a somewhat recent video of it doing this. Zip ahead to the 3:30 mark.

 

[Terminator857]

What race are we talking about?

How about race to solve driverless for significant percentage of the market in the U.S.? How about 1%? I think Waymo is at .000001% right now. How long before they reach 1%? I'll guestimate at least five years.

 

[diplomat33]

How about race to solve driverless for significant percentage of the market in the U.S.? How about 1%? I think Waymo is at .000001% right now. How long before they reach 1%? I'll guestimate at least five years.

Tesla is at 0% of the driverless market. So Waymo is still ahead of Tesla in that metric. I measure the metric based on actual deployment of driverless vehicles.

 

[Daniel in SD]

How about race to solve driverless for significant percentage of the market in the U.S.? How about 1%? I think Waymo is at .000001% right now. How long before they reach 1%? I'll guestimate at least five years.

There are about 3 million miles of paved road in the U.S. and .000001% of that is 160 feet.

 

[diplomat33]

There are about 3 million miles of paved road in the U.S. and .000001% of that is 160 feet.

Yeah, if you count all the roads in SF geofenced area, downtown Phoenix and Chandler geofenced areas, where Waymo has driverless, it is probably around 0.001%, still very low, but not as low as the 0.000001% @Terminator857 estimated.