- Joined
- Sep 1, 2024
- Messages
- 279
- Reaction score
- 199
- Location
- University Place, WA
- Cars
- 2022 Air Grand Touring
Pressing ignore button now...first time I've actually felt like I needed to use it.
LIDAR - Safest car on the road potential?
- Thread starter RL8
- Start date
What you're not getting is waymo uses server grade hardware. Terabytes of memory. Terabytes/s of memory bandwidth. Tens of thousands of tops. They use 360 spinning lidar, HD maps, etc.
These HD maps on their own are quite large so it is impossible to store an HD map of the U.S. on a consumer car.
This technology consumes 4000+ watts so unless you like having a car with 100 miles of range, no road tripping ability, and a battery that dies after only 6 hrs of continuous use, good luck.
The long standing argument by waymo and their engineers across reddit and other platforms is that the type of self-driving tesla and other consumer cars are pursuing is not possible. Camera only is a scapegoat for the fact that the type of driving tesla is pursuing is not possible for reasons other than being camera only.
If you remember from the early days of FSD you could cover up half of the cameras and the car would not throw any errors or drive any differently. What this seems to suggest is that there is a memory bandwidth limitation or some kind of compute limitation with processing a lot of sensors, and at any time you want to drive with the least amount of sensors possible. If you look at chinese cars with much better camera placement to see cross traffic, radar, lidar, etc. it seems to not provide any safety increase, which suggests they either are not able to effectively use them, or they lack compute to use them.
The only place they seem to have any benefit is in low speed parking where latency is not an issue due to speed.
- Joined
- Jan 4, 2022
- Messages
- 5,683
- Reaction score
- 10,936
- Location
- Santa Clarita, CA
- Cars
- 2026 Gravity Dream
- Gravity DE Number
- 68
- Referral Code
- AWNJLGKT
Every time one set of claims gets taken apart you don't respond to it, you just post a fresh set. The fatality claims from yesterday are gone, no acknowledgement, and now it's hardware. So let's do this one too, because almost nothing in it is right.
"HD maps are quite large so it is impossible to store an HD map of the U.S. on a consumer car."
You're posting this in a Lucid forum. Lucid's own Highway Assist has used high definition maps since the UX 2.4 update in 2024, and the hands-free driving that just rolled out to Gravity on DreamDrive 2 Pro only works on pre-mapped divided highways. GM's Super Cruise runs on lidar-scanned HD maps covering about 750,000 miles of North American roads, pushed to cars over the air. Mercedes Drive Pilot, the only certified Level 3 system you can buy in the US, uses lidar plus an HD map stored in backend data centres and updated constantly. Nobody stores a map of the entire country on the car. Maps are tiled and streamed for the roads you're on, which is how every one of these systems has worked for years. That's not an obscure fact. It's how the car in the OP's garage works.
"Waymo uses server grade hardware. Terabytes of memory. 4000+ watts. 100 miles of range, battery dies after 6 hours."
Waymo has never published memory, bandwidth, or wattage figures for its compute, so those numbers came from nowhere. What Waymo has published is that its 6th generation Driver cut sensor count by 42%, is expected to cost under $20,000 per unit, and is now running driverless across multiple cities. The 5th gen Jaguars have been in paid service for years running most of the day. If the compute drained the pack in six hours, the business would not exist.
"You could cover up half the cameras and the car would not throw any errors or drive any differently. What this suggests is a compute limitation."
What it actually suggests is the thing NHTSA is investigating. The open Engineering Analysis into FSD, covering 3.2 million vehicles, is specifically about the degradation detection system failing to notice when cameras are impaired by glare or fog and failing to warn the driver until right before the crash. You just described a known safety defect and presented it as evidence of clever engineering.
"The only place lidar seems to have any benefit is in low speed parking where latency is not an issue."
Backwards. Lidar is valued for long range detection at speed. Waymo's 6th gen suite sees out to 500 metres. Mercedes puts its lidar in the grille for highway Level 3, not parking lots. GM's entire hands-free domain is lidar-mapped highway. Low speed parking is where ultrasonics and cameras do fine. You've got the one use case exactly inverted.
"The long standing argument by waymo and their engineers across reddit is that consumer self-driving is not possible."
No source, again. Same as the Rivian cover-up, the Bilibili video, and "effectively zero at-fault fatalities."
You've been here 17 days, 43 posts, almost all of them about Tesla, and people in this thread are now hitting the ignore button. That's not because they can't handle a different opinion. It's because you keep stating things with total confidence that are checkable and wrong, and when they're checked you move on to the next batch.
Sources:
Mercedes-Benz world’s first automotive company to certify SAE Level 3 system for U.S. market | Automotive World
Mercedes‑Benz DRIVE PILOT self-certified in state of Nevada for SAE Level 3 conditionally automated driving
www.automotiveworld.com
Waymo begins fully autonomous ops with 6th-gen Driver, targets 1M weekly rides
Waymo is now running its 6th-generation Driver without safety drivers on public roads, marking the beginning of fully autonomous operations...
electrek.co
Meet the 6th-generation Waymo Driver: Optimized for costs, designed to handle more weather, and coming to riders faster than before
Waymo’s approach of designing both hardware and software from the ground up has been crucial to our success, and it continues to pay off as we introduce our sixth-generation hardware. We’ve significantly reduced the cost of our 6th-generation system while delivering even more resolution, range...
waymo.com
Tesla is one step away from having to recall FSD in NHTSA visibility crash probe
NHTSA escalated its Tesla FSD investigation to an Engineering Analysis covering 3.2M vehicles after finding the system fails to detect degraded visibility from glare and fog.
electrek.co
I posted the fatality claims and I believe it was removed for personal attacks by a moderator. I'm not going to explain again.
Lucid is not using HD maps at the same definition that are being used in a self driving car.
hd maps in self driving cars contain a lot more than just lane info. These would contain high density point clouds and they contain identifying information about every object that is being scanned, like whether it is a pothole, speed bump, etc. There is a lot more detail about how to drive on the roads vs. just a simple accurate lane map. There is definitely stuff I am forgetting.
Every object waymo scans with the point cloud has a "damage score" which is why waymo accidentally ran into that telephone pole in arizona. This would provide a level of safety as in a human could avoid a deadly accident (hitting a pedestrian) by crashing their car into something soft. Whereas maybe a tesla thinks every object is hard and this accident is unavoidable. This would also be an issue if you say swerve for a balloon when it is totally safe to hit.
There was an engineer on reddit who worked at cruise and he personally stated that the maps were very large and it's not possible to store an entire map of the U.S. on a car.
The way these maps are used is you scan the road with your 360 lidar and sync the output to the HD map. This gives you centimeter accurate position of where in a scene. Consumer GPS can be 20m or more off from the map which can cause you to miss a turn, think you are in a different intersection than you really are, think you are on the wrong road, etc.
The idea is at scale if you have a safety issue like the car can't identify a pothole or a speed bump, the HD map can take care of any scaling issues. The argument is this level of prior knowledge is required for safe, unsupervised driving. This has nothing to do with AI, E2E neural networks, etc. because those can still be used in a solution like this.
There was a video from a girl named Sophia Tung. She rode in a waymo for 6 hrs and had a full charged one (she lives near the depot) and the car started at 80% (79% since it lost 1% since she got in) and the car gave up after 6 hrs to return to the depot and recharge. It cuts off around 10%. She tried riding all day. There are also videos where the energy consumption screen is visible and it shows hilariously low ranges (like 30 miles remianing) with quite a large percentage of battery left. In the sophia tung test she went only 83 miles before the car gave up.
This no longer the case. FSD requires all the cameras to be visible. This literally hasn't been the case for at least 3 years. One of the limitations of FSD is that when the pillar camera on the sides see nothing (i.e. you're driving down a dark rural road at night and nothing is lit up), it thinks the camera is blinded and FSD gives an error. They have since solved this problem it seems.
I'm talking about chinese vehicles where lidar seems to do nothing for accident avoidance or reaction time.
If you use waymo there is evidence that every single speed bump that they drive over is premapped. If the speed bump is not premapped, the car won't slow down for it. There's evidence of waymos driving over speed bumps at hilariously incorrect speeds even though that are ridiculously visible.
If you don't have HD maps, you can't sync your lidar to the HD map and get this level of geolocation precision. You need this to accurately slow down for objects that you have issues with identifying. The argument with HD maps it provides a level of localized knowledge about everywhere you drive that you don't need to know everything about every road or intersection you encounter, you get a higher level of redundant safety in poor visibility scenarios (like snow/rain), and then basic driving behaviors become distilled to information that can be presented in map format. Also with GPS you can be 20m off from your actual location which can cause issues with thinking you're on a road you're not, accidentally being on the wrong side of the road, missing a turn at freeway speeds, thinking you're in the wrong intersection, etc.
The other issue is not having HD maps with this much driving information is that the performance does not scale to other countries. Waymo, for example is pretty easily able to handle RHD countries like england/japan/australia, whereas FSD performance is pretty bad there. Presumably also in europe it is much easier to tailor the driving to local regulations or to follow very specific rules like pickup and dropoff requirements at an airport.
The argument against not using HD mapping is it is not possible to achieve a scalable safety level without operating this way.
Huawei ads 5:
https://www.bilibili.com/video/BV1S...ck&vd_source=0e066e8af14fb39cf344050f887d347d
Chinese V13:
https://www.bilibili.com/video/BV1K...ck&vd_source=0e066e8af14fb39cf344050f887d347d
Cracked na v14.2:
https://www.bilibili.com/video/BV1PHpE6nEfd/?spm_id_from=333.1387.upload.video_card.click
Xpeng vla 6.3:
https://www.bilibili.com/video/BV1d...ck&vd_source=0e066e8af14fb39cf344050f887d347d
Look how bad these cars are performing.
Something people also don't know is mobileye own a pretty ridiculous and broad patent preventing cars from reading road signs.
Waymo mostly does not read road signs, and neither does tesla. Tesla seemed to have invented an alternative method to read speed limit signs, but I believe every other automaker on the road licenses mobileye's speed limit detection.
if you have an HD map this is much easier than a car without HD map to get around this limitation.
There are other details like tesla pushes out updates to FSD sometimes 3 or 4 times a week to test changes to the neural network. I believe china restricts update frequency and requires approval for new updates. This would hinder the speed you can iterate and fix issues.
Waymo mostly does not read road signs, and neither does tesla. Tesla seemed to have invented an alternative method to read speed limit signs, but I believe every other automaker on the road licenses mobileye's speed limit detection.
if you have an HD map this is much easier than a car without HD map to get around this limitation.
There are other details like tesla pushes out updates to FSD sometimes 3 or 4 times a week to test changes to the neural network. I believe china restricts update frequency and requires approval for new updates. This would hinder the speed you can iterate and fix issues.
For other information on the charging.
Waymo staging lot in austin uses natural gas charging with 40% efficiency. Puts their equivalent MPG around 19mpg and they charge at least 2x per day
Tesla uses unpaved lots with no charging, because they only need to charge every 2 or 3 days. They charge at superchargers overnight.
Waymo staging lot in austin uses natural gas charging with 40% efficiency. Puts their equivalent MPG around 19mpg and they charge at least 2x per day
Tesla uses unpaved lots with no charging, because they only need to charge every 2 or 3 days. They charge at superchargers overnight.