Tesla removes ultrasonic sensors from new Model 3/Y builds, soon Model S/X

[hridge2020]

After removing radar from their vehicles over the last year in a shift to their camera-based Autopilot system known as Tesla Vision, Tesla is now also removing ultrasonic sensors from new Model 3 and Model Y builds.

Tesla explains this change was made to coincide with the launch of the vision-based occupancy network, which allows them to replace the inputs generated by the sensors. Additional details of this occupancy network, which until now was limited to Full Self-Driving (FSD) Beta, were shared by Tesla’s Director of Autopilot software Ashok Elluswamy earlier this summer.

The ultrasonic sensors were primarily used to measure the distance to nearby objects using ultrasonic waves, like when parking your vehicle.

In an update to the Tesla Support page, the automaker says the removal of the sensors will begin in early October 2022 for Model 3 and Model Y built for North America, Europe, Middle East and Taiwan.

The change will take place globally over the next few months, before shifting to include the flagship Model S and Model X vehicles in 2023. (h/t: @Space_Taz)

If you are taking delivery soon, Tesla says you will be able to tell if your car has the sensors by looking for the little circles on the front and rear bumpers.

UPDATE 1:20pm PDT: Tesla has updated the Design Studio to no longer show the sensors on front and rear bumpers on all models, including the S/X.
While the transition occurs, Tesla says Park Assist, Autopark, Summon, and Smart Summon will be temporarily limited or inactive. All other features will remain fully functional.

Features*	Equipped with USS	Not equipped with USS
Forward Collision Warning	✓	✓
Automatic Emergency Braking	✓	✓
Lane Departure Warning / Avoidance	✓	✓
Emergency Lane Departure Avoidance	✓	✓
Pedal Misapplication Mitigation	✓	✓
Auto High Beam	✓	✓
Autowiper	✓	✓
Blind Spot Collision Warning Chime	✓	✓
AutoSteer†	✓	✓
Auto Lane Changes	✓	✓
Navigate on Autopilot	✓	✓
Traffic Light and Stop Sign Control	✓	✓
Park Assist	✓	Coming soon
Autopark	✓	Coming soon
Summon	✓	Coming soon
Smart Summon	✓	Coming soon

This is similar to how when Tesla removed radar, some features were limited or inactive until Tesla could fully verify their optimal operation within the fleet.

Tesla has improved but not yet fully restored those features, and they plan to follow the same pattern with this rollout.

“In the near future, once these features achieve performance parity to today’s vehicles, they will be restored via a series of over-the-air software updates. All other available Autopilot, Enhanced Autopilot and Full Self-Driving capability features will be active at delivery, depending on order configuration,” Tesla explains.

Source

Sponsored

 

[Crissa]

It's important to note while ultrasonics can see walls and cars and usually pedestrians and cyclists... They couldn't see posts or trailer hitches or whatnot.

I am wondering how they're going to deal with the 'object laying in front of the car' problem. The upper forward cameras can't see there.

-Crissa

 

[CyberGus]

Cybertruck will have a bumper camera, everyone else will buy additional insurance

 

[charliemagpie]

FSD has memory , and new occupancy network maps everything around it in 3D.

I'm guessing the car will monitor its surroundings at all times.. even when turned off.

It would take someone crawling out of a manhole and shimmy under the car out to the front and making sure their head stays below line of sight.

Then the robotaxi will drive over them

 

[JBee]

FSD has memory , and new occupancy network maps everything around it in 3D.

I'm guessing the car will monitor its surroundings at all times.. even when turned off.

It would take someone crawling out of a manhole and shimmy under the car out to the front and making sure their head stays below line of sight.

Then the robotaxi will drive over them

The difference between reactionary and predictive FSD.
Much better to know where things are before you get anywhere near them.
But they need a nose camera seperately IMHO, to look around corners and see straight in front of them in real time for unpredictable changes. Also more cameras means more redundancy for if one fails or gets dirty, especially for the primary direction of travel.

 

[charliemagpie]

The difference between reactionary and predictive FSD.
Much better to know where things are before you get anywhere near them.
But they need a nose camera seperately IMHO, to look around corners and see straight in front of them in real time for unpredictable changes. Also more cameras means more redundancy for if one fails or gets dirty, especially for the primary direction of travel.

FSD sees the environment, maps it and keeps track of it.

It is technically reactionary and predictive, in real time.

Once it maps its environment, when it travels, it recalibrates the position of every element relative to itself.

For example, when it parks near a curve, it will re-image the curve position within millimeters even though the camera doesn't see it.

Think about a condition where the car doesn't know exactly what's around it within say 5 millimeters.

 

[JBee]

FSD sees the environment, maps it and keeps track of it.

It is technically reactionary and predictive, in real time.

Once it maps its environment, when it travels, it recalibrates the position of every element relative to itself.

For example, when it parks near a curve, it will re-image the curve position within millimeters even though the camera doesn't see it.

Think about a condition where the car doesn't know exactly what's around it within say 5 millimeters.

The main reason why I said it like that is because the ultrasonics seem more reactionary in that they have to keep pinging a signal to give a give off some measurements whereas what vision sees gets plotted on the occupancy map and is retained for use at a later time. But maybe they plot ultrasonics as well? It just wouldn't have very good area resolution in comparison.

To predict anything it has to see it first, and not all things on the road are stationary, so having a nose cam gives it 1-2m more looking distance around corners and a backup for the windscreen ones get dirty.

 

[charliemagpie]

It doesn't have to ping anything. AI is good enough now to measure depth.

After that it can surmise its position to any object by mere advancement of Ai and brute computing power.

Back in early 1980's, my CEO was giving the Kmart CEO a tour of the computer center.. he asked me to tag along.. 'Hey Charlie, how fast is the 3090?'

FMD I thought, Who the fark knows that ? ,Jezus Gary.. throw me under a bus I thought !!!

3000 mips I said. I was right !
Talk about sweating it lol

So Elon mentions 144 trillion the other day. I think that was for 4 Dojo cabinets.. I'm not sure... Is that 40 odd billion times faster than the 3090 ?

In essence,.. the power of computing is beyond my comprehension. I think it was those NVIDIA 72 GPU's boxes which were superseded by Dojo?

I recall watching Jensen Huang's presentation ... was it a year ago ? .. those GPU boxes were 8 times faster than their previous GPU's.

4 Dojo cabinets, from my point of reference is 8*72 = 576 faster than the super fast GPUS of 2 years ago.


What is on the car is different, but it exceeds our traditional comprehension of practical capability. It is bordering magic.

 

[Crissa]

It would take someone crawling out of a manhole and shimmy under the car out to the front and making sure their head stays below line of sight.

FSD has been shown to forget things like people falling in front of it.

I certainly hope they have this solved, because yes, humans and animals walk, fall, in front of parked cars.

-Crissa

 

[JBee]

It doesn't have to ping anything. AI is good enough now to measure depth.

After that it can surmise its position to any object by mere advancement of Ai and brute computing power.

Back in early 1980's, my CEO was giving the Kmart CEO a tour of the computer center.. he asked me to tag along.. 'Hey Charlie, how fast is the 3090?'

FMD I thought, Who the fark knows that ? ,Jezus Gary.. throw me under a bus I thought !!!

3000 mips I said. I was right !
Talk about sweating it lol

So Elon mentions 144 trillion the other day. I think that was for 4 Dojo cabinets.. I'm not sure... Is that 40 odd billion times faster than the 3090 ?

In essence,.. the power of computing is beyond my comprehension. I think it was those NVIDIA 72 GPU's boxes which were superseded by Dojo?

I recall watching Jensen Huang's presentation ... was it a year ago ? .. those GPU boxes were 8 times faster than their previous GPU's.

4 Dojo cabinets, from my point of reference is 8*72 = 576 faster than the super fast GPUS of 2 years ago.


What is on the car is different, but it exceeds our traditional comprehension of practical capability. It is bordering magic.

You still can't compute what you can't see first. Imgaine a car turns into a main road from around a corner the camera couldn’t see. Or a car that is parked on the side behind a truck and pulls out. There's physically no way you can predict that with computation.

 

[firsttruck]

FSD has been shown to forget things like people falling in front of it.

I certainly hope they have this solved, because yes, humans and animals walk, fall, in front of parked cars.

-Crissa

or even sleep in front or in back of parked vehicles. Also young children that are not old enough yet to learn the dangers of being close to cars

 

[firsttruck]

You still can't compute what you can't see first. Imgaine a car turns into a main road from around a corner the camera couldn’t see. Or a car that is parked on the side behind a truck and pulls out. There's physically no way you can predict that with computation.

Yup, all the situations and many more occur. These should not be treated as corner cases. Everyday people die or are injured in similar situations.

Need corner cameras to better see cross traffic.

I hope Tesla takes the saving from removing the 12 ultrasonics and use that saving to offset the cost of 4 corner mounted cameras.

 

[JBee]

Might not matter so much if they're exactly on the corners, insomuch they overlap in such a way there are no deadspots on either side, that could of been captured from somewhere on the vehicle. It already has cameras on the tail and sides but is still missing a camera on the front, the windscreen one is mounted nearly halfway back on the car so it's more like a middle camera.

 

[firsttruck]

Might not matter so much if they're exactly on the corners, insomuch they overlap in such a way there are no deadspots on either side, that could of been captured from somewhere on the vehicle. It already has cameras on the tail and sides but is still missing a camera on the front, the windscreen one is mounted nearly halfway back on the car so it's more like a middle camera.

No, I think it matters a lot that the new cameras be at the corners of the vehicle.

The Tesla diagram that shows the cameras & all the overlapping FOVs is misleading. Those FOVs are only realized IF there are NO fixed or moving objects in the environment that are blocking the view paths of cross traffic. Things like other vehicles in adjacent lanes, trees, fences, walls of building, car/trucks parked along the sides of roadway, crowd of people standing/waiting on the sidewalk at the corner, and many more things.

It is very common that B-pillar camera will not be able to see high-speed cross traffic at sufficient distance because of the B-pillar camera's view being blocked. The higher the speed of cross traffic the further down the cross street the B-pillar needs to see but that also means there is higher chance there will be objects in the environment that block the view. Having higher res cameras can not help here because of the obstructions there is nothing for the camera to see unless it has superman like x-ray vision.

I think these vision obstruction problems will occur much more frequently in east coast cities and even suburbs because things are much close together than places like Los Angeles, Dallas, etc that are more spread out.

Sponsored