WSJ story/cautionary tale on self driving/Driver + comparo

[JJE]

The WSJ had a pretty disturbing story today on the limitations of Tesla's camera -based self driving tech. For those who subscribe to the WSJ this is the link:

The Hidden Autopilot Data That Reveals Why Teslas Crash

https://www.wsj.com/articles/the-hidden-autopilot-data-that-reveals-why-teslas-crash-334cccab

This got my attention bc of two incidents this weekend when the Rivian Driver + system suddenly almost drove me off the road. Conditions were clear and the road well- marked and I can't see what triggered the car to behave randomly. I have generally been happy with the Driver + system but it made me question my confidence. If I hadn't been on the ball and attentive both times we would definitely have crashed.

Unlike Tesla, Driver + uses radar and ultrasound in addition to cameras and in principle has more redundancy. But my takeaway from this weekend is not to be complacent.

I asked ChatGPT4 to compare the Tesla and Rivian systems. I know this has been discussed extensively. FWIW I am sharing the summary from Chat here (I recognize the irony in relying on another form of machine learning):

Rivian and Tesla both use advanced driver-assistance systems (ADAS) in their vehicles, but there are notable differences in their approaches:

Sensor Suite:

Tesla: Tesla primarily relies on a camera-based system, known as Tesla Vision, along with ultrasonic sensors. They have phased out radar from their vehicles, arguing that vision-based systems can provide better data for the neural networks to process, resulting in safer and more effective self-driving capabilities.

Rivian: Rivian uses a combination of cameras, radar, and ultrasonic sensors. This multi-modal approach provides a more comprehensive view of the environment, combining the strengths of each type of sensor. Cameras are excellent for detecting and classifying objects, radar is useful for detecting objects in poor visibility conditions, and ultrasonic sensors help with close-range detection.Software and Processing:

Tesla: Tesla's Autopilot and Full Self-Driving (FSD) systems use neural networks and continuous over-the-air software updates to improve their capabilities. Tesla vehicles collect data from the entire fleet, which is used to train and refine their neural networks.

Rivian: Rivian's ADAS system, known as "Rivian Driver+", also uses machine learning and over-the-air updates. While Rivian's system may not have the same level of fleet data as Tesla's, it still employs a robust suite of sensors and processing to enable features like lane keeping, adaptive cruise control, and highway assist

.Philosophy and Approach:Tesla: Tesla has a more aggressive approach toward achieving full autonomy, often releasing features in a beta state to collect data and improve functionality. Their goal is to reach full self-driving capability, where the car can drive without human intervention.

Rivian: Rivian appears to take a more cautious approach, focusing on driver assistance rather than full autonomy. Their system is designed to assist drivers and improve safety rather than replace human drivers entirely.Overall, Rivian's use of a combination of cameras, radar, and ultrasonic sensors provides a different set of capabilities compared to Tesla's camera-centric approach. Each company’s system reflects its philosophy and strategy towards the future of autonomous driving.

Sponsored

 

[Treebeard]

The WSJ had a pretty disturbing story today on the limitations of Tesla's camera -based self driving tech. For those who subscribe to the WSJ this is the link:

The Hidden Autopilot Data That Reveals Why Teslas Crash

https://www.wsj.com/articles/the-hidden-autopilot-data-that-reveals-why-teslas-crash-334cccab

This got my attention bc of two incidents this weekend when the Rivian Driver + system suddenly almost drove me off the road. Conditions were clear and the road well- marked and I can't see what triggered the car to behave randomly. I have generally been happy with the Driver + system but it made me question my confidence. If I hadn't been on the ball and attentive both times we would definitely have crashed.

Unlike Tesla, Driver + uses radar and ultrasound in addition to cameras and in principle has more redundancy. But my takeaway from this weekend is not to be complacent.

I asked ChatGPT4 to compare the Tesla and Rivian systems. I know this has been discussed extensively. FWIW I am sharing the summary from Chat here (I recognize the irony in relying on another form of machine learning):

Rivian and Tesla both use advanced driver-assistance systems (ADAS) in their vehicles, but there are notable differences in their approaches:

Sensor Suite:

Tesla: Tesla primarily relies on a camera-based system, known as Tesla Vision, along with ultrasonic sensors. They have phased out radar from their vehicles, arguing that vision-based systems can provide better data for the neural networks to process, resulting in safer and more effective self-driving capabilities.

Rivian: Rivian uses a combination of cameras, radar, and ultrasonic sensors. This multi-modal approach provides a more comprehensive view of the environment, combining the strengths of each type of sensor. Cameras are excellent for detecting and classifying objects, radar is useful for detecting objects in poor visibility conditions, and ultrasonic sensors help with close-range detection.Software and Processing:

Tesla: Tesla's Autopilot and Full Self-Driving (FSD) systems use neural networks and continuous over-the-air software updates to improve their capabilities. Tesla vehicles collect data from the entire fleet, which is used to train and refine their neural networks.

Rivian: Rivian's ADAS system, known as "Rivian Driver+", also uses machine learning and over-the-air updates. While Rivian's system may not have the same level of fleet data as Tesla's, it still employs a robust suite of sensors and processing to enable features like lane keeping, adaptive cruise control, and highway assist

.Philosophy and Approach:Tesla: Tesla has a more aggressive approach toward achieving full autonomy, often releasing features in a beta state to collect data and improve functionality. Their goal is to reach full self-driving capability, where the car can drive without human intervention.

Rivian: Rivian appears to take a more cautious approach, focusing on driver assistance rather than full autonomy. Their system is designed to assist drivers and improve safety rather than replace human drivers entirely.Overall, Rivian's use of a combination of cameras, radar, and ultrasonic sensors provides a different set of capabilities compared to Tesla's camera-centric approach. Each company’s system reflects its philosophy and strategy towards the future of autonomous driving.
[/QUOTE
Day one of driving the truck decided it wanted to rapidly decelerate in traffic at 75 mph. I was ready and hit the juice pedal. Never have trusted computers. Lucky it was not heavy traffic.

 

[Electrified Outdoors]

Lately I have noticed phantom braking is more of an issue on our R1 vehicles. Those familiar with Tesla will be very familiar with the phantom braking issue.

The Rivian system hasn't been that way but it seems to be happening more lately.

 

[Rivianready]

I think that in the long run, Tesla’s reliance on cameras will be a flawed strategy. Rivian’s hardware, while better, will take some time to mature. I have noticed a number of bugs using Driver+ including phantom braking and late braking. it is definitely not ready for prime time, so keep your eyes on the road and hands on the wheel!

 

[Noplacelikeloam]

I use it often but am now cautious and dont trust it TBH. It recently tried to veer left sharply while at 70mph on a highway. That was an "oh sh1t" moment I do not want to repeat.

 

[JJE]

That is exactly what happened to me. I had no problems until the latest sw update btw.

 

[RexRemus]

I have had a few phantom braking moments - most often seem tied to approaching an overpass, not sure why, but that seems to be when I've noticed the events, I suspect it might have something to do with the shadow being cast on the road surface (depending on time of day and direction of travel). And I have had a handful of "swerves" but I can't say that at any moment in either case I felt in great danger of any kind. I cover the pedal, and if the vehicle shows, I am on the pedal to correct it. I am always hands on the wheel - if I feel a sudden direction shift, I counter steer (almost always disengages D+) and nothing "scary" happens.

Is there a gap between sensing those things and reacting? Of course. but I do think it's on the driver to NEVER "trust" ANY driver assistance. They tell you to stay focused and be aware - I'm pretty sure they mean it. I still find that at least on all the routes I've used it on D+ is generally pretty good. I have gripes - a lot of them - but I still find it worth using because it does at least reduce some mental burden for general lane centering or minor changes in speed, etc. I still think it's a good thing when it works - and it mostly does. But I never, ever, "trust" it. I am always ready to "drive". I never take my foot away from the pedal, or hands away from the wheel.

I can't speak for anyone else, but I've never had an experience where I felt a steering correction was "just about to slam me into a barrier" or another vehicle - maybe it WOULD, if I were not hands on wheel, and not closely watching things - easy enough to slip into a distracted state. And if I had to sorta "snap to attention" and THEN correct, I bet it would feel a lot more scary.So my advice is - don't do that. Use the tech, it's there to be used, I enjoy seeing HOW it responds to certain things, I treat it like beta testing because it IS. "Oh it was a little weird how it handled that curve" "That was odd, it shouldn't have followed that lane marker" - but I can make those observations because I was already ready to intervene and catch them right away.

I guess all I'm trying to say is - I find it odd that anyone comments about "not trusting" these systems because it seems obvious to me that we should NEVER trust them. But I am also an advocate for using them - they can be great, a vast majority of the time, but all of them are far from perfect and they WILL mess up - assume that, and be prepared for it. Ultimately you are the driver and YOU need to be in control of the vehicle. But don't get twisted about "broken trust" - there should be no trust

 

[JJE]

The issue is that some of the car's movements when the system malfunctions are quite abrupt and startling. One ends up reacting both to the overall circumstances on the road, and to the car's erroneous reaction.

Depending on the cognitive load on the driver it could be too much for short term memory to manage both in time to avoid an accident. Dark, rainy slick nights at speed on unfamiliar roads etc combined with a sharp, uncommanded movement out of your lane is no joke.

I also find the system on balance to be a plus. The system is pretty conservative and tends to shut off if there is any confusion. But there are dangers that simply being attentive may not always be enough to overcome.

 

[NY_Rob]

I have had a few phantom braking moments - most often seem tied to approaching an overpass, not sure why, but that seems to be when I've noticed the events, I suspect it might have something to do with the shadow being cast on the road surface (depending on time of day and direction of travel).

I've had the same thing happen a couple of time now, and each time it was approaching a large overpass that was casting a shadow down on to the roadway preceding the overpass. It seems that Driver+ thinks that shadow is a solid object on the roadway ahead of us?

 

[CrazyOne]

Self driving will require Lidar or a new one yet to be invented . They can train the neural networks for various scenarios, but they can only do so much with flawed sensors.

 

[RexRemus]

I've had the same thing happen a couple of time now, and each time it was approaching a large overpass that was casting a shadow down on to the roadway preceding the overpass. It seems like Driver+ thinks that shadow is a solid object on the roadway ahead of us?

That's my assumption as well, somehow it's reading the shadow/light change as an "object" of some kind obstructing the road - might just be an issue with exposure/dynamic range of the front camera - where it dials things way down for the sun, but that causes the shadowy parts to appear very dark, and in those few moments before it can adjust/adapt the exposure it looks nearly solid black - and so...there must be something there "blocking" the lane/lines

 

[RivianRunner]

Self driving will require Lidar or a new one yet to be invented . They can train the neural networks for various scenarios, but they can only do so much with flawed sensors.

Humans can drive with only two eyes, no radar, no lidar and no high-precision mapping. Why do you think AI won't be superior with more "eyes" and no lapses in attention?

Scratch that, there are humans who have over a million accident and injury free miles with only one eye, and no other sensors. The driving environment is designed to work with vision only.

Sponsored