Waymo One launches

[diplomat33]

Just saw that some Arizona media folks followed some Waymos for 170 miles and captured some clips:

That's a very informative video. Thanks for sharing. I think it shows that while Waymo's self-driving is impressive, it is certainly not Level 5 yet.

 

[CarlK]

Earlier you criticized me that I don't understand software. Now I think you are the one who doesn't get it. NN is trained to reach a well specified behavior. It can't do more than what the target is. It can do a task better but it can't outsmart humans. Maybe 100 years from now on quantum computers.

You are right you know nothing about how AI NN works. You set a reward system but not a specific behavior of how it should go. Using mongo's analogy you don't tell it to hit like Mike Trout or JD Martinez. You reward it every time it hits better. Eventually it will hit better than anyone ever able to by practicing 100's or 1000's times a human players ever could.

 

[Bladerskb]

Just a note Elon never uses Lx to describe the FSD capability. Those definitions are too broad and can't characterize usefulness and difficulties among each individual system. No two systems are the same whether one puts them on the same level or not. This of course also could indicate different thinkings between Elon and others of how this is approahed.

 

[Bladerskb]

Regarding overal plan. I think we can agree that Tesla's development of AP has had a few reboots on terms of approach.
My take on the current setup (based mostly on the Karpathy talk) is that Tesla is doing something like this:
1. Aquire large compote cluster for NN training and testing
2. Put all data, labels training images, test cases, NN shape guidance in their repository
3. Retrain entire NN on every check in, training against every source, validating against every test case
4. Testing includes running them against labeled real world data at silicon speed
5. Possibly putting a bunch of copies in a simulated driving environment for further testing

This setup would be why Elon says no one (that he knows of) is working on the level of generalized solution that Tesla is because Tesla has only one NN/ code base. It is not the normal SW route of adding features as seperate modules.

To add stop signs, for instance, they would take the current code base which trains a NN that handles NoA and such, then add in the stop sign positive and negative reference images along with test case and rebuild the entire thing. Based on the results, the training data and parameters get updated to correct the new behavior and address any issues generated in the original feature set.

As each feature is added, each previous feature gets better due to the increased library of things it 'understands' and things it is tested against (now that you know what a stop sign is, ignore stop signs not on your section of the road)

Net result is that progress comes in big chunks (when all tests pass), and that we won't see the big step change until the 3.0 HW is installed.

You are clearly misinformed on what you are talking about.
You ppl need to stop being easily lead astray by elons flattery.

Waymo's ChauffeurNet

Learning to Drive: Beyond Pure Imitation – Waymo – Medium

Comments by CEO of Voyage

Oliver Cameron on Twitter

Full Paper

[1812.03079] ChauffeurNet: Learning to Drive by Imitating the Best and Synthesizing the Worst

 

[mongo]

You are clearly misinformed on what you are talking about.
You ppl need to stop being easily lead astray by elons flattery.

So confused, I stated that Tesla likely restarted their development efforts and gave my interpretation of Karpathy's SW 2.0 talk.
What did any of that have to do with Elon beyond his statement of a general solution (which has nothing to do with flattety)?

 

[strangecosmos]

So confused, I stated that Tesla likely restarted their development efforts and gave my interpretation of Karpathy's SW 2.0 talk.
What did any of that have to do with Elon beyond his statement of a general solution (which has nothing to do with flattety)?

Some people are unwilling or unable (or a bit of both) to have a constructive, open-minded, good faith, friendly discussion. When someone responds to your technical conjecture with personal insults or insinuations, there is no productive way to continue the conversation. It is an unfortunate drawback of the unmoderated culture on this subforum.

I have a much more pleasant experience of TMC now that I’ve ignored everyone who is unfriendly, although I still see these half-conversations in most active threads and I can tell what’s going on.

There is also a seemingly insurmountable communication barrier with at least once user. It may be a language barrier issue. I’ve tried to clarify the meaning of phrases like “in 2019” (it of course means anytime from January 1, 2019 to December 31, 2019), but to no avail. If someone misunderstands certain English phrases, and doesn’t pay any attention to/doesn’t accept explanations of them, how are you supposed to have a conversation? It’s crazy-making to try.

So my advice is don’t waste your good heart and mind on this futile, painful task. We are deep in troll country, and need to avoid the swamps.

 

[acoste]

You are right you know nothing about how AI NN works. You set a reward system but not a specific behavior of how it should go. Using mongo's analogy you don't tell it to hit like Mike Trout or JD Martinez. You reward it every time it hits better. Eventually it will hit better than anyone ever able to by practicing 100's or 1000's times a human players ever could.

Haha, hard to make a conversation with you if you say I'm wrong, then you repeat what I wrote ...


ps: when I said specific behavior, I meant desired output, not the internals.

 

[CarlK]

Haha, hard to make a conversation with you if you say I'm wrong, then you repeat what I wrote ...


ps: when I said specific behavior, I meant desired output, not the internals.

Nice try but you're still wrong big time. You said machine could not do better than the target and it can't out smart human. You absolutely have no idea how it works. Did you even know AlphaGo beat world's number one ranked Go player with program written by programmers who are probably at best mediocre players.

 

[Bladerskb]

So confused, I stated that Tesla likely restarted their development efforts and gave my interpretation of Karpathy's SW 2.0 talk.
What did any of that have to do with Elon beyond his statement of a general solution (which has nothing to do with flattety)?

You literally said

This setup would be why Elon says no one (that he knows of) is working on the level of generalized solution that Tesla is because Tesla has only one NN/ code base. It is not the normal SW route of adding features as seperate modules.

The idea that Tesla has a approach that hasn't been researched when in actuality has been researched and shown to be flawed. As proven AGAIN by the recent Waymo research papers on pure end to end imitation learning using millions of driving data and abstract imitation learning using unlimited simulation data.

The idea that Tesla's development is

1. Collect data
2. Label data
3. Retrain and Deploy NN
4. ???
5. Profit!

 

[electronblue]

Some people are unwilling or unable (or a bit of both) to have a constructive, open-minded, good faith, friendly discussion.

I have not watched this long but to me it seems to be you with your passive aggressiveness being the worst offender and first to quit open-minded, good-faith discussions when they get difficult for you. May I suggest you purge your ”ignore list” and stop advocating putting people on it — very unfriendly act in itself that — and get back to talking to all of us. Will make for less an echo chamber all around.

As a friendly bonus it would also make for less duplicate stuff you post of things others are already discussing at length.

Regarding overal plan. I think we can agree that Tesla's development of AP has had a few reboots on terms of approach.
My take on the current setup (based mostly on the Karpathy talk) is that Tesla is doing something like this:
1. Aquire large compote cluster for NN training and testing
2. Put all data, labels training images, test cases, NN shape guidance in their repository
3. Retrain entire NN on every check in, training against every source, validating against every test case
4. Testing includes running them against labeled real world data at silicon speed
5. Possibly putting a bunch of copies in a simulated driving environment for further testing

Well put thesis thank you. Do you believe Tesla is abandoning the multiple NNs they have used currently and instead going for one single NN for everything?

 

[mongo]

You literally said


The idea that Tesla has a approach that hasn't been researched when in actuality has been researched and shown to be flawed. As proven AGAIN by the recent Waymo research papers on pure end to end imitation learning using millions of driving data and abstract imitation learning using unlimited simulation data.

The idea that Tesla's development is

1. Collect data
2. Label data
3. Retrain and Deploy NN
4. ???
5. Profit!

Yeah, I know literally said:

This setup would be why Elon says no one (that he knows of) is working on the level of generalized solution that Tesla is because Tesla has only one NN/ code base.

Hence saying:

What did any of that have to do with Elon beyond his statement of a general solution (which has nothing to do with flattety)?

I didn't say it hasn't been researched, not have I made a claim as to its effectiveness. However, if previous research indicates it doesn't work, then that would support Tesla being the only one doing it (not necessarily in a good way).

More clarification in next post...

 

[mongo]

Well put thesis thank you. Do you believe Tesla is abandoning the multiple NNs they have used currently and instead going for one single NN for everything?

I think when Karpathy joined, they restarted the NN side of thing. The training data was likely still usable in some form.
My guess: the NN exists as one object. However, the internals of the NN (lane detection, sign detection, vehicle gap adjusting) could still be tested (and thus trained) individually. What the suspected SW 2.0 approach removes is having separate groups working on separate aspects with separate NNs and then trying to merge them at the end.

General can also mean that they are using some more typical, research paper supported, methods, but they do everything over the complete data set. Whereas some people may start at a neighborhood, then region, then city, then country, then another country, then worldwide with changes each time, Tesla could be going for all or nothing.

Analogy: training a robot to walk, progressive: flat, high mu ground, bumpy ground, hill, ruts, mountains, ice. General: all topology cases at once.

Similar to the Lidar debate: Lidar gets to a working system sooner, but Tesla is going for pure vision, which Lidar system will need also for completeness. So, in terms of apparent progress, Lidar system start out ahead, but then run into many of the same problems.

End results: the Tesla approach may look poor until it works, then it looks good (perhaps great).

Just my thoughts.

 

[CarlK]

LIDAR provides some more info that camera does not. The end result is the NN trained from that would not be as smart.

Similar to the Lidar debate: Lidar gets to a working system sooner, but Tesla is going for pure vision, which Lidar system will need also for completeness. So, in terms of apparent progress, Lidar system start out ahead, but then run into many of the same problems.

End results: the Tesla approach may look poor until it works, then it looks good (perhaps great).

Just my thoughts.

Absolutely agree. LIDAR provides some extra info that camera does not. That will make things seem to be easier in the beginning but the end result is the NN trained from that will not be as smart as that trained by camera only. That's why Elon says using LIDAR is like on crutches. You will need to relearn how to walk when you take crutches away.

The only argument is whether camera alone is enough to get us there. Imo it is definitely enough. Modern camera has already exceeded capability of human eye in pretty much every measure. No reason to think it could not do the job.

 

[mongo]

LIDAR provides some more info that camera does not. The end result is the NN trained from that would not be as smart.


Absolutely agree. LIDAR provides some extra info that camera does not. That will make things seem to be easier in the beginning but the end result is the NN trained from that will not be as smart as that trained by camera only. That's why Elon says using LIDAR is like on crutches. You will need to relearn how to walk when you take crutches away.

The only argument is whether camera alone is enough to get us there. Imo it is definitely enough. Modern camera has already exceeded capability of human eye in pretty much every measure. No reason to think it could not do the job.

Agree, although I would say the NN without Lidar needs to do more procesing, but the system as a whole could be equally as 'smart'.

 

[heltok]

LIDAR provides some more info that camera does not. The end result is the NN trained from that would not be as smart.

If they use the Waymo network architecture, the high level NN will use very similar input and output. The main difference is that the bounding boxes will be slightly larger due to larger distance and size uncertainty, the other cars will drive more noisily and some cars and some on road obstacle might be missed. But on the other hand a much larger dataset will be collected with more rare dangerous scenarios trained on.

Would not be surprised if this is how Karpathy et al were planning to do this.

Tesla seems to be using unreal3d as a simulation environment: Autopilot simulation!

I would recommend watching Zoox videos as they also use unreal3d:

(23:14 and forward)

 

[ricebucket]

That's a very informative video. Thanks for sharing. I think it shows that while Waymo's self-driving is impressive, it is certainly not Level 5 yet.

You're confusing level 5 with human-like. The video showed several instances where the car did not react like humans (except for the left turn scenario, but it is not clear that the driver is actually turning the wheel vs holding the wheel). I don't see a problem with not behaving like humans if the car was able to successfully avoid any obstacles and bring the passenger to the destination, without assistance. Just this is the definition of level 5 driving.

And reacting like humans will matter less and less as more autonomous vehicles appear on the road. They should not be constrained by human behavior. (For example, there's no point in asking a car to use feet to move right? Why should it make turns like a human?)

 

[ricebucket]

LIDAR provides some more info that camera does not. The end result is the NN trained from that would not be as smart.


Absolutely agree. LIDAR provides some extra info that camera does not. That will make things seem to be easier in the beginning but the end result is the NN trained from that will not be as smart as that trained by camera only. That's why Elon says using LIDAR is like on crutches. You will need to relearn how to walk when you take crutches away.

The only argument is whether camera alone is enough to get us there. Imo it is definitely enough. Modern camera has already exceeded capability of human eye in pretty much every measure. No reason to think it could not do the job.

You're missing the forest for the trees. If lidar can help (which you admit in your statement), then why not have it? More information is always better. So what if cameras are better than humans? If you can make an even better car with lidar, then why not build one? A bicycle can run faster than my legs, but that didn't stop people from building cars that run faster than bikes.

Yes, cost is an issue, but that's a business problem, not a technical one. (And LIDAR costs are bound to come down.)

By the way, "NN trained from that will not be as smart as that trained by camera only" sentence does not make sense.

 

[diplomat33]

You're confusing level 5 with human-like. The video showed several instances where the car did not react like humans (except for the left turn scenario, but it is not clear that the driver is actually turning the wheel vs holding the wheel). I don't see a problem with not behaving like humans if the car was able to successfully avoid any obstacles and bring the passenger to the destination, without assistance. Just this is the definition of level 5 driving.

And reacting like humans will matter less and less as more autonomous vehicles appear on the road. They should not be constrained by human behavior. (For example, there's no point in asking a car to use feet to move right? Why should it make turns like a human?)

I am not talking about the car not behaving like a human. I am talking about disengagements, specifically when the safety driver had to take control to do a maneuver. When a human has to take control, by definition, that's not Level 5. That is what I was talking about.

 

[ricebucket]

I am not talking about the car not behaving like a human. I am talking about disengagements, specifically when the safety driver had to take control to do a maneuver. When a human has to take control, by definition, that's not Level 5. That is what I was talking about.

Did we watch the same video? There's no confirmed disengagements in that video. The reporters were outside the car and they couldn't have known about any disengagements.

 

[diplomat33]

Did we watch the same video? There's no confirmed disengagements in that video. The reporters were outside the car and they couldn't have known about any disengagements.

Maybe it was a different video I watched but I remember a reference to the safety driver having to sometimes take control to make left turns into the Waymo parking lot area.