WHY Tesla Removed Radar from Model 3 & Y (Sr. AI Director)

Andrej Karpathy recently spoke at CVPR 2021, a virtual conference on Computer Vision and Pattern Recognition and he shared a MULTITUDE of USEFUL insights revealing the HOW and the WHY Tesla is going ALL IN ON VISION for their Full Self Driving Software.

Support Cleanerwatt by:

*** All video and pictures are used with permission or in accordance to the copyright owner's stated policies and use allowance. ***

Image & Video Clip Sources:

1. Tesla Video/Pictures: Courtesy of Tesla, Inc.
2. Waymo Robotaxi Video: Courtesy of Waymo
3. Mobileye Robotaxi Video: Courtesy of Mobileye

Data Sources:

DISCLAIMER & DISCLOSURE: All of the content found in this video is based on my own opinion and should not be regarded as financial advice. I am not a financial advisor and this is NOT in any way a recommendation or offer to buy or sell securities. While the information in this video is believed to be correct at the time of recording, no guarantees are being made about the accuracy of the information presented in the video. Please consult a licensed professional, or do your own research before making any financial decisions. As of the recording of this video, Jonathan (host) is NOT invested in Tesla (TSLA).

[asalaices]

What about the added safety that radar provides while driving in fog and rain? Isn't that an added bonus?

[christianwilhite]

Just got a new M3 without radar. I can say phantom braking is still a thing.

[rhtufts]

Stopped vid when presenter felt the need to explain a crutch analogy to me.

[iPigee]

I have a vision only model Y. Phantom braking is present and rampant at times. Need to check your facts.

[LysanderLH]

I did not buy a Tesla to beta test their deadly software issues eg. Phantom Braking, Emergency Steering correction etc, which causes the car to slam on its brakes and steer away from what it perceives to be a threat. There is no override for these issues and each time I get into my car, I disable ALL automated driving functions and I never use adaptive cruise or Autopilot because they are dangerous! So that’s most of the functionality of the car made useless.
This is a moral issue. Tesla are STILL knowingly selling cars with potentially lethal software issues which owners cannot override! Why aren’t governments insisting on mass product recalls?

[richardsnyder]

Those rainy conditions you showed are what it’s like every day in Florida during the summer

[royhornyak]

Humans don't drive using sight alone, we use severa of our senses besides sight, touch, hearing, even balance. A camera can not duplicate this alone. Also Radar is of special use when heavy rain causes spray on the roads, a camera can be blinded by this, also how about heavy fog? I really think that Tesla are making a mistake removing the radar, even if it is just a backup system.

[maxmiller]

Interesting video!!!
You should try to make videos on Investment...

[pihi]

I've been saying this since the radar cancellation news: it's a sensor fusion problem. This might be a limit of current machine learning, but still Karpathy is doing some PR in his speech. You don't and can't get the same data from visual light camera as you do from a radar. Radar gives you direct doppler measurements of speed and since it uses different wavelengths, it gives you different data (not obstructed by water in the air as much). I get it, it is possible to drive with vision only (humans can do it), so it's certainly possible for an AI. BUT, saying additional data radar gives you is useless or even 'causing noise' is just a statement "we can't handle the data correctly" spoken as a PR person, not a scientist. I don't like that. I would like to see comparison of speed from vision vs speed from radar. I would like to see what kind of noise they are getting; is it sensor data-level noise, sensor fusion level noise, or some higher level NN noise. i would like to see why they are having presumable problems with labelling radar data correctly (mapping radar on top of vision). Is the radar too poor in spatial accuracy to be usable? Would a slightly better radar do it? Radar's extreme precision measuring speed is the reason why police are using them to measure traffic instead of relying on their visual assessment of speed. Or why Arecibo was a killer instrument, untouchable by anything else, in detecting asteroid speeds and positions and even obtaining surface images. Is Karpathy saying, in his professional opinion, that radar data is also useless in astronomy? I don't like that kind of talk from a researcher/development leader. These things Elon can say and be forgiven, but not Karpathy.

[santiagodraco]

"This seems to make full sense doesn't it?" (narrator referring to radar + vision and "doubling down" on vision when they disagree). No, it doesn't make full sense. It does if you want to believe Tesla I

The issue, if there really is one, is that radar + vision is a software coding issue. The idea that radar is holding you back is nonsense. All you need to do is develop your AI algorithms to utilize radar in situations where the system knows that vision cannot provide sufficiently accurate data and you have to fall back to radar for safety. Such as in a rainstorm, or heavy fog, or snow.

[chrisheath]

Very clear presentation, thanks. One thing that is fundamental to dropping radar, that Andrej talked about, was the fact that, if a car in front brakes heavily, radar will actually 'lose' the car - the signal disappears, and reappears as the radar locks back on. Radar needs a lot of pre-processing to derive a "correct" signal, and the return signal is weak. If the "noise" ( think distorted signal) is close to the "signal", it becomes impossible to have a useful signal. The other issue with radar is that other cars may also have radar, so you need processing to correctly identify your own signal. These problems can probably be overcome, but is it worth pursuing, when video cameras have over 100 times the information. ( I'm sure people would love to hear from an expert on finer details, of why the radar was dropped - it has horrified some people who think the cars will crash more, because they've dropped radar!)

[srinyacharya]

A lot of people have complained that the vision only approach has led to significant increase in "phantom braking". Hopefully they will be able to resolve this with incremental OTA software updates.

[cooksey]

I swear over half this video was repeating himself and saying they are going with Visio and getting rid of radar.. we get it bro they are getting rid of radar 🙄

[jginart]

We just got a new long range model Y with no radar this month. On our first road trip, we have suffered 4 phantom breaking incidents on the highway on autopilot. I hope they fix this.

[andyrechenberg]

5:11 I have first-hand experience that removing radar does not eliminate phantom braking. The reason for the phantom braking that I experienced wasn't due to radar and vision disagreeing because my Model Y is one of the new ones without radar. I have experienced phantom braking about 5 times in ~400 miles of highway driving. Hopefully the car will get better over time. Thanks for the video. Cheers.

[deanmcmanis]

Admittedly, we don't absolutely need to use all of our senses to get around in the world, and many people could (and do) easily survive having just sight, or even just hearing. But the truth is that having multiple senses helps us perceive the world, and our surroundings, better and more accurately. Some senses are indeed distracting, and I wouldn't want a smell or taste sensor added to help my car drive better. Although sound might help. The value of using multiple sensors is in the areas where each sensor has issues. For cameras, poor contrast/glare, and occlusions (where the camera view is blocked or otherwise impaired) render the cameras useless. Now for a blocked camera, you do have multiple cameras available, just as long as the field of view is overlapping. As for the argument that an automated driving system only needs to be as good as people, then why would you use infrared sensors for being able to see better at night? I think that Tesla's decision is more based on cost and component availability. Tesla's research in their own battery development was partially done because Asian corporations have long dominated battery materials and production, worldwide. And Tesla's vertical integration has given them more flexibility and autonomy. Likewise, Tesla's move to Pure Vision has probably been accelerated by the recent crippling chip and vendor part shortages that have left many thousands of otherwise completed vehicles unfinished and unshipped, despite growing demand and sales. Also if Tesla can establish themselves with a proven Autopilot system that doesn't need pricey sensors, then they can make a strong business case to sell their Autopilot software worldwide to many other automakers as a profit center.

[nickbailey]

I've had my Tesla for 2 weeks now. No radar. Phantom breaking is still a thing.

[chungloankenny]

I want to see video about why Tesla didn't remove radar from S and X

[lazyfrogonalog]

To me the removal of radar is an obvious extension to Tesla getting rid of processing single images in favour of processing an all round "video" image. The software is easier to write, quicker to react with less noise. This was a great decision that will progress them much faster to full self driving.

[jeffbransky]

A human can drive safely, although slowly, on a snow covered country road. They do in all the time in the rural Northeast. Tesla Vision can’t see the edges of the road so it will give up and shut down.