7 ms·

I'm sure that Tesla will have to eventually bring back radar into the stack. The problem with a vision system only is that it relies on visual clues that can be

I'm sure that Tesla will have to eventually bring back radar into the stack. The problem with a vision system only is that it relies on visual clues that can be easily blinded by the conditions of the road. For instance a sun glare, high beams from traffic, water on the windshield, or simply a small depression on the road terrain. These are the same things that can cause a traffic accident for a human with human eyes, so there's really no logical reason to assume that you can create a more performing and safer computer driver by emulating human vision only. Radar solves all these problems by enhancing the perception of the car and creating redundancy so the car can cross-check its assumptions about the road, the terrain, and the obstacles with two different and independent perception systems. I can't understand why they removed radar. You don't need to be an expert to understand why vision alone won't work.

HWR_14 4y ago
> I can't understand why they removed radar. The cost per unit times 500k+ units/year is significant. Also, I think the radar they were using started getting deprecated by the manufacturer. Now, do I think they should have included RADAR -or- stopped selling FSD? Yes. I don't see how they can deliver FSD without RADAR.
- stingrae 4y ago
  I was under the impression, that they removed it because they were having a hard time sourcing the components in the current chip shortage.
  - JoshCole 4y ago
    False impression and common. We don't need to guess here. There was an engineer talk at the Computer Vision and Pattern Matching conference in 2021 in the Workshop on Autonomous Driving by Andrej Karpathy senior director of AI at Tesla. In the latter portions of the talk he dives deep into why they choose to go all in on video with compelling empirical results showing it to be far more accurate than vision + radar. https://news.ycombinator.com/item?id=31414374 https://news.ycombinator.com/item?id=31414374 They do end up doing sensor fusion, but not when you would think. At training time they use radar still, because at training time you can do things to correct for radars errors, for example, you can use hindsight to reason about what happened. At inference time, the uncertainty of hallucinating radars is a real risk to safety. It says things aren't there when they are. It says things are there when they aren't. That makes it a lot harder for the car to plan out the next actions and promotes very dangerous action - like slamming on the brakes - which can surprise other drivers and cause accidents.
ggreer 4y ago
Radar has poor resolution, which means it can't always differentiate between an object near the car's path and an object in the car's path. The radar can give false positives when there's a truck on the side of the road or a metal sign on an overpass. Then you have to use the vision cameras to figure out what's really going on. Radar also loses its "lock" sometimes, which can be an issue with cars braking quickly. The radar will say "no obstacles detected" when the car in front of you is slamming on the brakes. Again, you have to use the vision cameras to figure out what's actually going on. The people at Tesla decided that the radar was creating more noise than it was helping with, so they removed radar last year. Andrej Karpathy explains it in far more detail in this talk: https://www.youtube.com/watch?v=g6bOwQdCJrc&t=7m https://www.youtube.com/watch?v=g6bOwQdCJrc&t=7m
- ra7 4y ago
  These are Tesla-only problems because they were using Continental radars from 2012. Everyone else has developed advanced radars like Waymo with their high-resolution imaging radar that can track objects that are moving and stationary [1] or see through fog [2]. This is another way of saying Tesla wanted to cut costs and not invest in radar development. [1] https://blog.waymo.com/2020/03/introducing-5th-generation-waymo-driver.html https://blog.waymo.com/2020/03/introducing-5th-generation-wa... [2] https://blog.waymo.com/2021/11/a-fog-blog.html https://blog.waymo.com/2021/11/a-fog-blog.html
  - JoshCole 4y ago
    If they were Tesla only problems, why did other companies feel the need to advance the state of the art in their radar hardware? They did so because poor radar equipment wasn't a Tesla only problem, but was a problem that many identified. Poor radar was not a Tesla only problem, but a common problem. Now the world state advanced to a world in which WayMo had more advanced radars and Tesla did not have more advanced radars, but instead stopped using bad radars. Both companies however made the decision to abandon the radars that they previously had. Both identified getting rid of their former radars as the correct solution to the problem of radars which did not meet their needs. So it isn't the case that it was a Tesla only problem and it isn't the case that the two companies had terribly different ideas about how to solve the problem. Both companies improved their car by relying on technology to make up for shortcoming in radar systems. The difference was in how they did it. Tesla did it by devoting more effort to the vision stack whereas WayMo did it by devoting more effort to hardware improvements. In the video you dismiss there are segments that show Tesla's vision-only system successfully identifying distances despite the presence of fog. They also identify things that are stationary. They also identify things that are moving. So the things you cite for supporting your thesis that Tesla failed aren't examples of things that Tesla is doing that WayMo isn't. They are examples of things that both cars are doing. That means they are worthless for making a claim of superiority. Having failed to establish superiority you move on to claiming that Tesla wanted to cut costs and that this was their motivation for doing this. This is cast in a bad light as if cutting cost harms safety. In reality training data is precious and collecting it improves performance. Cutting cost allows greater volume of production by lowering the capital cost of production. This results in more sales. This results in more data. The net result is a better training set, improving performance. You aren't even correct that cutting costs actually results in less spending on solving the critical problems: cutting costs leads to increased sales which leads to increased revenue and in turn increased profits which leads to more capital for development, not less capital for development, which means more capital gets allocated to solving self-driving problems. As such cutting costs has the impact of increasing the amount that is spent on researching the problem as time goes on. The big difference in their choices isn't that Tesla is approaching it in a bad way and WayMo in a good way. It is that WayMo is getting funding from projects that aren't itself. It can afford to make bad decisions in terms of cost-effectiveness when it thinks that it improves the probability of eventual success. This sort of spending model has historically been simultaneously underwhelming and overwhelming in what it produced. Overwhelming because it does things like get us to the moon, but underwhelming in that it does so in such a cost ineffective way such that we don't return. This is shaping up to be the difference between WayMo and Tesla in practice. WayMo handles driving in fenced off locations well! This is the overwhelming awesome thing it does. However, at the same time, it has extremely low production volume and doesn't support all cities. This is like how we got to the moon, but then didn't get to keep going there. Meanwhile, Tesla has literally millions upon millions more self-driving miles. However, this actually works out in practice to WayMo harming public health in comparison to Tesla? Why? Well, self-driving plus humans is already better than humans driving alone. WayMo has technology that could save lives to the tune of some fraction of the ninety people in the US who die every day in accidents not dying. WayMo isn't letting that technology save those lives. Instead, Tesla is letting a similar technology save those lives. As Tesla scales it captures more of the market, it stops more of those deaths. Thus in terms of their net benefit in terms of death reduction, WayMo actually loses by far. Every moment they don't scale while Tesla does scale, they lose some more. The only hope for them catching up is to achieve greater scale than Tesla which is going to give them a cost effectiveness problem or alternatively they need Tesla to mess up in a truly massive way that forces regulators hands - something that regulators and Tesla don't want at present, because the current trend is lives being saved from human errors.
Fronzie 4y ago
The motivations I have seen were: - If radar/lidar fails, vision needs to work. If vision works, radar/lidar are not needed. - Sensor fusion is hard: outliers in radar data ended up hurting more than the non-outliers helped. The second one sounds to me like the real reason: Sensor fusion is not trivial and if it doesn't work correctly, you'll get the worst of all worlds.
- PeterisP 4y ago
  If radar/lidar fails, it's perfectly fine for vision to work just barely enough to slow down for a few seconds while telling you that you're going to have to self-drive yourself.
mandevil 4y ago
Back in 2005-6 I was working on autonomous off-road vehicles for the US Army, part of the far outer fringes of the [Future Combat System] (https://en.wikipedia.org/wiki/Future_Combat_Systems https://en.wikipedia.org/wiki/Future_Combat_Systems). In our tests we found that MMWR had major speed limitations even with a leader-follower system (where the manned leader built terrain models as it was manually driven and shipped them to the followers which then tried to terrain-match the leader's MMWR images). We figured based on sensor range versus safe-stopping acceleration rates it was limited to about 30kph. We also had tremendous problems with dust etc. up in the air causing the radar to think it was a wall and panic stop. And further problems, since we were off-road, the radar LOVED puddles and any sort of standing water- it looked perfectly flat and so wanted to go there every time. Various attempts were made to compensate (some people tried sensor fusion with visual and IR cameras, which doing in real time is hugely hard, our team tried to do it by focusing on human-robot collaboration, but we never found a good way to adjudicate differences between human and robot, and if the human is always having to intervene anyway, the robot's not giving you much help). None of them were working when FCS was cancelled and our project focus shifted.
- aaroninsf 4y ago
  Thank you for sharing this real-world experience! It is precisely these sorts of really-very-hard challenges that have led me to conclude that "FSD" absent general AI is not plausible, unless you narrow the definition of FSD to mean something that can only be used in the 80% case. The thing that perplexes me is why the industry isn't chasing the 80% case, e.g. follow-the-leader cross country traffic on interstates... I always think, the interstate system is tailor made in so many dimensions (regulatory and physical) for roll out of limited but robust automation, and the number of highway miles there probably represent the vast majority of total miles driven... In shipping it seems you could be predicting and real-time sharing local conditions to allow for graceful exit from automation when required, and, avoid the situations where it was reasonably well. There's always the outliers, though... tumbleweeds, smoke, fog, bad actors...
  - mandevil 4y ago
    Yeah, back then I also thought that a leader-follower convoying system would be implemented first (with local drivers for the last mile problem), because that system made sense and was straightforward- with good logistics software (to make sure that the leader is driving the farthest distance or connecting multiple leaders together) you could easily save 2/3 or more of your driver costs. Then a few years ago I talked to an actual software consultant to shipping companies, and asked him why none of that had happened in over a decade. He felt was that the companies were too fragmented for that. Most shipping companies are small companies, where basically the owners wife keeps a list of all of their shipping loads and destinations. Maybe Walmart has the access to capital to acquire that technology and the scale to profit from it, but he felt that no one else could. Which was why he left that field and went into industrial robotics, where the business case can actually close.
nemothekid 4y ago
RADAR would not have helped in this situation. Even if the oncoming car wasn't there, the Tesla still made an illegal maneuverer - in no state I can think of is it illegal to cross 2 double yellow lines. Either 1.) The camera's fidelity isn't good enough to make out the double yellow lines 2.) A software bug caused the Tesla to drive to aggressively.
jumelles 4y ago
> I can't understand why they removed radar. Elon Musk's ego.
nouveaux 4y ago
Radar and lidar may be needed but this is not a good example of it. I am watching this video on youtube recorded with a camera. I can clearly see the lanes in this video and I can clearly see the oncoming traffic. The car should not have crossed over and it should be able to detect the large light approaching in the other lane. This is a completely failure of the AI that radar and lidar would fail to solve.
johnwalkr 4y ago
Aside from the visual clues you mentioned, I like to use the example of a screen door or vertical blinds. Most people have the experience of inadvertently focusing on these things a bit wrong and "feeling" that the objects are much closer or farther than reality. Computer vision can make the same mistakes easily, and in fact can often perform much worse than human vision. Backlit scenes are difficult, frontlit scenes are more difficult, and scenes with a lot of repetition and few large features are almost impossible. Examples of almost impossible scenes are backlit beach on a sunny day and cloudy day while skiing, and all kinds of conditions while driving, such as frontlit or sidelit with a wet road.
lifeplusplus 4y ago
Only if visual system came with some fish memory and common sense. AV systems seem to be very focused on now data but lack mental model and decision making. They should really develop for no markings or lanes environment first
JoshCole 4y ago
> These are the same things that can cause a traffic accident for a human with human eyes, so there's really no logical reason to assume that you can create a more performing and safer computer driver by emulating human vision only. Okay. This is a concrete claim that I think is falsifiable. I'll give not just one, but multiple arguments. Argument 1: Connect Four, Chess, Checkers, Go, Limit Hold'em, No Limit Hold'em, Poker, DOTA, and Starcraft II are all examples of games in which AI outperforms even the best humans despite having equal sensory access. Driving can be reduced to a problem comparable to playing an imperfect information game and we can play imperfect information games at superhuman levels using computers. Therefore, it is reasonable to expect that even human-level sensory access can enable safer than human performance. Argument 2: Currently we have statistics on driver-assistance programs that demonstrate that human + technology assistance is superior to human alone. This holds true even in cases in which those assistance technologies are only using vision. So we have a contradiction. If it wasn't possible to do better than human with only human senses, we wouldn't be able to point to many different situations in which driving with an AI clearly outperforms humans, but we can. For at least those subsets of AI driving, we can say that an AI driver will outperform humans. > I can't understand why they removed radar. Check out https://www.youtube.com/watch?v=g6bOwQdCJrc https://www.youtube.com/watch?v=g6bOwQdCJrc for a very detailed breakdown of why they did it.

HWR_14 4y ago

> I can't understand why they removed radar. The cost per unit times 500k+ units/year is significant. Also, I think the radar they were using started getting deprecated by the manufacturer. Now, do I think they should have included RADAR -or- stopped selling FSD? Yes. I don't see how they can deliver FSD without RADAR.

stingrae 4y ago
I was under the impression, that they removed it because they were having a hard time sourcing the components in the current chip shortage.
- JoshCole 4y ago
  False impression and common. We don't need to guess here. There was an engineer talk at the Computer Vision and Pattern Matching conference in 2021 in the Workshop on Autonomous Driving by Andrej Karpathy senior director of AI at Tesla. In the latter portions of the talk he dives deep into why they choose to go all in on video with compelling empirical results showing it to be far more accurate than vision + radar. https://news.ycombinator.com/item?id=31414374 https://news.ycombinator.com/item?id=31414374 They do end up doing sensor fusion, but not when you would think. At training time they use radar still, because at training time you can do things to correct for radars errors, for example, you can use hindsight to reason about what happened. At inference time, the uncertainty of hallucinating radars is a real risk to safety. It says things aren't there when they are. It says things are there when they aren't. That makes it a lot harder for the car to plan out the next actions and promotes very dangerous action - like slamming on the brakes - which can surprise other drivers and cause accidents.

ggreer 4y ago

Radar has poor resolution, which means it can't always differentiate between an object near the car's path and an object in the car's path. The radar can give false positives when there's a truck on the side of the road or a metal sign on an overpass. Then you have to use the vision cameras to figure out what's really going on. Radar also loses its "lock" sometimes, which can be an issue with cars braking quickly. The radar will say "no obstacles detected" when the car in front of you is slamming on the brakes. Again, you have to use the vision cameras to figure out what's actually going on. The people at Tesla decided that the radar was creating more noise than it was helping with, so they removed radar last year. Andrej Karpathy explains it in far more detail in this talk: https://www.youtube.com/watch?v=g6bOwQdCJrc&t=7m https://www.youtube.com/watch?v=g6bOwQdCJrc&t=7m

ra7 4y ago
These are Tesla-only problems because they were using Continental radars from 2012. Everyone else has developed advanced radars like Waymo with their high-resolution imaging radar that can track objects that are moving and stationary [1] or see through fog [2]. This is another way of saying Tesla wanted to cut costs and not invest in radar development. [1] https://blog.waymo.com/2020/03/introducing-5th-generation-waymo-driver.html https://blog.waymo.com/2020/03/introducing-5th-generation-wa... [2] https://blog.waymo.com/2021/11/a-fog-blog.html https://blog.waymo.com/2021/11/a-fog-blog.html
- JoshCole 4y ago
  If they were Tesla only problems, why did other companies feel the need to advance the state of the art in their radar hardware? They did so because poor radar equipment wasn't a Tesla only problem, but was a problem that many identified. Poor radar was not a Tesla only problem, but a common problem. Now the world state advanced to a world in which WayMo had more advanced radars and Tesla did not have more advanced radars, but instead stopped using bad radars. Both companies however made the decision to abandon the radars that they previously had. Both identified getting rid of their former radars as the correct solution to the problem of radars which did not meet their needs. So it isn't the case that it was a Tesla only problem and it isn't the case that the two companies had terribly different ideas about how to solve the problem. Both companies improved their car by relying on technology to make up for shortcoming in radar systems. The difference was in how they did it. Tesla did it by devoting more effort to the vision stack whereas WayMo did it by devoting more effort to hardware improvements. In the video you dismiss there are segments that show Tesla's vision-only system successfully identifying distances despite the presence of fog. They also identify things that are stationary. They also identify things that are moving. So the things you cite for supporting your thesis that Tesla failed aren't examples of things that Tesla is doing that WayMo isn't. They are examples of things that both cars are doing. That means they are worthless for making a claim of superiority. Having failed to establish superiority you move on to claiming that Tesla wanted to cut costs and that this was their motivation for doing this. This is cast in a bad light as if cutting cost harms safety. In reality training data is precious and collecting it improves performance. Cutting cost allows greater volume of production by lowering the capital cost of production. This results in more sales. This results in more data. The net result is a better training set, improving performance. You aren't even correct that cutting costs actually results in less spending on solving the critical problems: cutting costs leads to increased sales which leads to increased revenue and in turn increased profits which leads to more capital for development, not less capital for development, which means more capital gets allocated to solving self-driving problems. As such cutting costs has the impact of increasing the amount that is spent on researching the problem as time goes on. The big difference in their choices isn't that Tesla is approaching it in a bad way and WayMo in a good way. It is that WayMo is getting funding from projects that aren't itself. It can afford to make bad decisions in terms of cost-effectiveness when it thinks that it improves the probability of eventual success. This sort of spending model has historically been simultaneously underwhelming and overwhelming in what it produced. Overwhelming because it does things like get us to the moon, but underwhelming in that it does so in such a cost ineffective way such that we don't return. This is shaping up to be the difference between WayMo and Tesla in practice. WayMo handles driving in fenced off locations well! This is the overwhelming awesome thing it does. However, at the same time, it has extremely low production volume and doesn't support all cities. This is like how we got to the moon, but then didn't get to keep going there. Meanwhile, Tesla has literally millions upon millions more self-driving miles. However, this actually works out in practice to WayMo harming public health in comparison to Tesla? Why? Well, self-driving plus humans is already better than humans driving alone. WayMo has technology that could save lives to the tune of some fraction of the ninety people in the US who die every day in accidents not dying. WayMo isn't letting that technology save those lives. Instead, Tesla is letting a similar technology save those lives. As Tesla scales it captures more of the market, it stops more of those deaths. Thus in terms of their net benefit in terms of death reduction, WayMo actually loses by far. Every moment they don't scale while Tesla does scale, they lose some more. The only hope for them catching up is to achieve greater scale than Tesla which is going to give them a cost effectiveness problem or alternatively they need Tesla to mess up in a truly massive way that forces regulators hands - something that regulators and Tesla don't want at present, because the current trend is lives being saved from human errors.

Fronzie 4y ago

The motivations I have seen were: - If radar/lidar fails, vision needs to work. If vision works, radar/lidar are not needed. - Sensor fusion is hard: outliers in radar data ended up hurting more than the non-outliers helped. The second one sounds to me like the real reason: Sensor fusion is not trivial and if it doesn't work correctly, you'll get the worst of all worlds.

PeterisP 4y ago
If radar/lidar fails, it's perfectly fine for vision to work just barely enough to slow down for a few seconds while telling you that you're going to have to self-drive yourself.

mandevil 4y ago

Back in 2005-6 I was working on autonomous off-road vehicles for the US Army, part of the far outer fringes of the [Future Combat System] (https://en.wikipedia.org/wiki/Future_Combat_Systems https://en.wikipedia.org/wiki/Future_Combat_Systems). In our tests we found that MMWR had major speed limitations even with a leader-follower system (where the manned leader built terrain models as it was manually driven and shipped them to the followers which then tried to terrain-match the leader's MMWR images). We figured based on sensor range versus safe-stopping acceleration rates it was limited to about 30kph. We also had tremendous problems with dust etc. up in the air causing the radar to think it was a wall and panic stop. And further problems, since we were off-road, the radar LOVED puddles and any sort of standing water- it looked perfectly flat and so wanted to go there every time. Various attempts were made to compensate (some people tried sensor fusion with visual and IR cameras, which doing in real time is hugely hard, our team tried to do it by focusing on human-robot collaboration, but we never found a good way to adjudicate differences between human and robot, and if the human is always having to intervene anyway, the robot's not giving you much help). None of them were working when FCS was cancelled and our project focus shifted.

aaroninsf 4y ago
Thank you for sharing this real-world experience! It is precisely these sorts of really-very-hard challenges that have led me to conclude that "FSD" absent general AI is not plausible, unless you narrow the definition of FSD to mean something that can only be used in the 80% case. The thing that perplexes me is why the industry isn't chasing the 80% case, e.g. follow-the-leader cross country traffic on interstates... I always think, the interstate system is tailor made in so many dimensions (regulatory and physical) for roll out of limited but robust automation, and the number of highway miles there probably represent the vast majority of total miles driven... In shipping it seems you could be predicting and real-time sharing local conditions to allow for graceful exit from automation when required, and, avoid the situations where it was reasonably well. There's always the outliers, though... tumbleweeds, smoke, fog, bad actors...
- mandevil 4y ago
  Yeah, back then I also thought that a leader-follower convoying system would be implemented first (with local drivers for the last mile problem), because that system made sense and was straightforward- with good logistics software (to make sure that the leader is driving the farthest distance or connecting multiple leaders together) you could easily save 2/3 or more of your driver costs. Then a few years ago I talked to an actual software consultant to shipping companies, and asked him why none of that had happened in over a decade. He felt was that the companies were too fragmented for that. Most shipping companies are small companies, where basically the owners wife keeps a list of all of their shipping loads and destinations. Maybe Walmart has the access to capital to acquire that technology and the scale to profit from it, but he felt that no one else could. Which was why he left that field and went into industrial robotics, where the business case can actually close.

nemothekid 4y ago

RADAR would not have helped in this situation. Even if the oncoming car wasn't there, the Tesla still made an illegal maneuverer - in no state I can think of is it illegal to cross 2 double yellow lines. Either 1.) The camera's fidelity isn't good enough to make out the double yellow lines 2.) A software bug caused the Tesla to drive to aggressively.

jumelles 4y ago

> I can't understand why they removed radar. Elon Musk's ego.

nouveaux 4y ago

Radar and lidar may be needed but this is not a good example of it. I am watching this video on youtube recorded with a camera. I can clearly see the lanes in this video and I can clearly see the oncoming traffic. The car should not have crossed over and it should be able to detect the large light approaching in the other lane. This is a completely failure of the AI that radar and lidar would fail to solve.

johnwalkr 4y ago

Aside from the visual clues you mentioned, I like to use the example of a screen door or vertical blinds. Most people have the experience of inadvertently focusing on these things a bit wrong and "feeling" that the objects are much closer or farther than reality. Computer vision can make the same mistakes easily, and in fact can often perform much worse than human vision. Backlit scenes are difficult, frontlit scenes are more difficult, and scenes with a lot of repetition and few large features are almost impossible. Examples of almost impossible scenes are backlit beach on a sunny day and cloudy day while skiing, and all kinds of conditions while driving, such as frontlit or sidelit with a wet road.

lifeplusplus 4y ago

Only if visual system came with some fish memory and common sense. AV systems seem to be very focused on now data but lack mental model and decision making. They should really develop for no markings or lanes environment first

JoshCole 4y ago

> These are the same things that can cause a traffic accident for a human with human eyes, so there's really no logical reason to assume that you can create a more performing and safer computer driver by emulating human vision only. Okay. This is a concrete claim that I think is falsifiable. I'll give not just one, but multiple arguments. Argument 1: Connect Four, Chess, Checkers, Go, Limit Hold'em, No Limit Hold'em, Poker, DOTA, and Starcraft II are all examples of games in which AI outperforms even the best humans despite having equal sensory access. Driving can be reduced to a problem comparable to playing an imperfect information game and we can play imperfect information games at superhuman levels using computers. Therefore, it is reasonable to expect that even human-level sensory access can enable safer than human performance. Argument 2: Currently we have statistics on driver-assistance programs that demonstrate that human + technology assistance is superior to human alone. This holds true even in cases in which those assistance technologies are only using vision. So we have a contradiction. If it wasn't possible to do better than human with only human senses, we wouldn't be able to point to many different situations in which driving with an AI clearly outperforms humans, but we can. For at least those subsets of AI driving, we can say that an AI driver will outperform humans. > I can't understand why they removed radar. Check out https://www.youtube.com/watch?v=g6bOwQdCJrc https://www.youtube.com/watch?v=g6bOwQdCJrc for a very detailed breakdown of why they did it.