Waymo will beat Tesla FSD because of LIDAR

[cyberos]

*Disclaimer: I work for Google, but not Waymo
*Overall opinion: I use the Cybertruck's FSD all the time, even to drive 2-minutes down the street; it is awesome because it's here NOW

After seeing many FSD threads this past week and enjoying the Cybertruck's FSD myself I wanted to share a brief crash course on autonomous driving

References:

• MIT's course on Artificial Intelligence: Implications for Business Strategy
• Lecture, "Understanding Robotics"
• Lecture transcripts attached for convenience

Robotics 101: AI in the Physical World, with Sensors and Actuators
At a high level robots are just automated devices that perform physical tasks in the real world using 1. sensors (inputs) and 2. actuators (outputs)

Tesla is Level 3 Autonomy, Waymo is Level 4

All Self Driving Cars "Map" Between Feature Space (virtual) and Physical Space (real)

• All of the sensors (cameras, LIDAR, radar, microphones, etc.) are used to create a map of our real world in a virtual world called "feature space"
• The car "thinks" and "acts" entirely in this feature space
• It needs the sensors to keep populating the feature space with objects, to try to mimic our physical world as much as possible
• This why the Cybertruck "drives over curbs" while Waymo "sees the curbs", Waymo's feature space is better populated, closer to our real world

Tesla FSD is Parallel Autonomy (human-in-the-loop), Waymo is Series Autonomy (no human, mostly)
In robotics there are two types of autonomy, series and parallel:

• Series autonomy (Waymo): either human in control or vehicle in control, not both
• Parallel autonomy (Tesla FSD): aka "Supervised" FSD

Tesla's Big Failure: No LIDAR
In one sentence, robots need "laser imaging, detection and ranging" (LIDAR) to accurately perceive depth, occlusion, etc.

Longer answer is that LIDAR facilitates building a 3D point map of the world. Initially cameras were meant only for object detection. Tesla admirably expanded camera-based computer vision into FSD, but the lack of LIDAR is still problematic

Tesla's Reason for Camera-Only Computer Vision FSD: Cost
LIDAR (and the now removed radar from early model Teslas) is expensive and it's not just the hardware. The Cybertruck FSD update boasts that its neural network / machine learning model replaces over "300K lines of explicit C++ code". That's a lot of software engineering hours and millions of dollars saved

My (unverified) opinion is that there are nowhere near as many LIDAR 3D map examples as there are computer vision image and video examples. This means training a LIDAR machine learning model is not (yet) feasible

Waymo cleverly overcame this limitation by transforming the LIDAR 3D point map into a more manageable format. I have not had the time to read all the research, but Waymo solved the LIDAR point map problem, where a mess of 3D dots inhibited using machine learning fully

All of Waymo's research came at an extreme cost and they had Google / Alphabet money. Tesla just couldn't afford to run at a loss and had no other massively profitable business units to draw from

Sponsored

 

[TickTock]

I think the argument against LIDAR is humans don't have it and yet can driver better than Waymo (or Tesla). My experience with AR on the very low cost Quest headset is vision easily gets within a couple centimeters which should be more than enough for driving a car. There are definitely cases where I wish I had LIDAR feeding my brain - for example when I've turned left in front of a car far away at night only to discover it was actually a Jeep much closer than I thought. Vision based FSD will have similar difficulties in this scenario. However, I've never caused a wreck - just had to goose it a bit - something Tesla's are very good at.

 

[sean-techventures]

Like the previous post, Tesla's uses of photogrametry based FSD is already better than humans at parsing out the real world (e.g most humans can determine depth just like Tesla's solution). Lidar only adds noise to the model that it has to interpet.

 

[sean-techventures]

And Waymo is just in SF? lol get real dude

 

[cyberos]

Same thought for me initially, and now it’s in four cities and growing

 

[Outdoors]

Same thought for me initially, and now it’s in four cities and growing

What is the growth rate? I think the number of cities is growing. Yet if it adds cities at current rate. I will see it in my town 2039.

Thanks for the TLDR. The course was not brief, and it did crash as I couldn't bear to read it all.

Will you ask your bosses at Google when I can buy a Waymo thing. The answer will be. Never.

 

[M0unt41nm4n]

Great post by the OP. I will agree Waymo with LIDAR will "beat" Tesla's visual-only, but I will disagree with anything that hints that Tesla won't get as good as humans or better, and likely get pretty close to Waymo's LIDAR in the future. I do AI development for a living as well. LIDAR is most certainly helpful as an additional data point for FSD AI. The more features (as we call them - different data points) you have, the better AI gets in the models. That can mean that Waymo's use of LIDAR makes it potentially better since it adds a ton more features. However, multiple cameras will provide depth, similar to how your eyes work by providing triangulated vision. Notice if you close one, eye, you lose distance measuring to a strong degree. But when you use two eyes, you can get back the sense of depth. I am not familiar with Tesla's video/cameras, but as long as they have multiple cameras, they should be able to get to depth and reasonably be able to see curbs, bumps, obstacles, etc, and gather range information. FSD just needs a lot more training, which we are providing for them using our FSD. Here is a snippet about AI for people who do not do it which is taught in Machine Leanring 101: In order to get a good model you need one or two items (or both). You need more training data (volume), or more features, or both. Tesla can get there... they just need a lot more time to gather data and push it through thier models - which is why they ask us when there is an issue. They can mark it as an abberation and push it through retraining their models We are lucky to be a part of this.

So in a nutshell, the OP is right, it may may not get to perfection because it loses the LIDAR feature data. But over massive amounts of training, it will get darned close and likely be better than a human, which is really all we care about. The advantage Tesla has over Waymo is there are a lot more Teslas in the world and will have a ton more data than Waymo to perfect its models.

 

[BornToFly]

Wrong. The problem is not with vision. Cameras are just fine. The problem is the "intelligence" and with deciding how to drive and handle what is happening in the world around the car. Waymo mostly gets around this with very highly mapped strict geofencing, which is why it doesn't work in the rest of the world. Tesla will win with its new AI training center and billions of miles of training data.

 

[StainlessVoyager]

I dont think we have exact info, but I would assume Waymo is easily over 100K per car. Financially FSD makes a lot more sense. if self driving is going to be a real world widely adopted thing cost has to make sense.

 

[M0unt41nm4n]

Wrong. The problem is not with vision. Cameras are just fine. The problem is the "intelligence" and with deciding how to drive and handle what is happening in the world around the car. Waymo mostly gets around this with very highly mapped strict geofencing, which is why it doesn't work in the rest of the world. Tesla will win with its new AI training center and billions of miles of training data.

I'm not sure who you are replying to... but yeah, its not "overall" due to vision. Its due to features input to the model. LIDAR + Vision gives it a crap ton more input features than just vision alone. As I stated earlier, the more features you have, the better the model will be. Yes, you are correct, as more data from Tesla goes into the model, the better it gets, which is what I was saying. But its a hard sell against LIDAR + Vision. Long term it will get really close. But LIDAR+ Vision will always have an edge just from the feature set. That said, as I stated earlier, its not going to matter because after a lot of data, it will beat a human and the LIDAR will be negligible. But it will be better. OP is correct on this.

 

[Gaximus]

A car that has only vision, with enough training data, and processing power, can drive as good as humans(but better because it never loses focus, it can evaluate a 1000+ a second) and never get tired or fatigued). Humans hit curbs all the time. But, with radar, it could be better in some situations, fog, rain, dark.
I hate that Tesla removed the radar, if you’re going to have a robot do human tasks, why limit it to human capabilities (Tesla factory robots aren’t limited to lifting a couple hundred pounds a few times a day)

Very thorough post, BTW.

 

[lrpena]

humans don't have lidar and drive perfectly well a camera based system should be able to do the same or even better since it has 360 field of view. .

 

[Amuma]

I truly feel Tesla will also implement some form of Lidar in combo with vision in their fully autonomous vehicles.. Vision will not work in all conditions ie rainy/snowy conditions.

 

[Deleted member 17810]

Like the previous post, Tesla's uses of photogrametry based FSD is already better than humans at parsing out the real world (e.g most humans can determine depth just like Tesla's solution). Lidar only adds noise to the model that it has to interpet.

"it adds noise" is something musk mentioned during his presentation a couple years ago, because it was slowing down development and was too costly (computing wise) to implement with visual only.

That "noise" is what prevents an autonomous system from dropping out an vehicle behind a sign, or a stationary box truck, or an emergency vehicle on the side of the road, and actually seeing it nearly instantaneously.

Lidar actually imitates a "sense" that visual computation can't. hidden distance tracking.

They haven't proven it yet, they've talked about it before, but it was too slow. Maybe on robo day.

humans don't have lidar and drive perfectly well a camera based system should be able to do the same or even better since it has 360 field of view. .

Do you know why sobriety tests have a touch your nose with your eyes closed test?

Humans can track distances very acutely. It's the entire premise of modern sports.

 

[Cybergirl]

I wish someone would drop a Waymo car off in front of my house and instruct it to take me to anywhere. Make it easy. Take me to the Jewel Foods store a couple blocks away. It won't even get to the end of the block if it moves at all. Waymo cars have absolutely no understanding of the physical driveable world - none. They are sophisticated warehouse robots engineered to operate in city-sized, tightly constrained warehouses. It's nonsensical to claim that Waymo will beat Tesla because of Lidar. Even in a constrained operating environment, Waymo's problem isn't so much engineering as it is economic. Nothing succeeds in this world if it violates the "laws" of economics, and Waymo is extremely secretive about costs. In my opinion, Waymo will fail because they've created a complex system the cost of which increases linearly with the number of cars deployed, i.e, it is not economically scalable given the higher hardware, support, and maintenance costs compared to a future Tesla robotaxi network.

Sponsored