Lidar vs Camera revisited

[Bladerskb]

Thats the whole point - HD map vs Map.

No I described in both posts, an HD map. (Moderator Edit)

Post #1:

For a map to be correct it needs to have a high refresh rate.
For a map to be able to be refreshed automatically, it needs to be localize-able.
Meaning it has to have a level of detail that allows you to link different elements of the road together and also
the level of detail for you to be able to spot what has changed, regardless if its a huge change or a very small change.

Post #2 which summarizes what i said in #1:

crowdsourced detailed map that is automatically updated

 

[johnmodely]

@Bladerskb you understand that tesla's whole approach of why they don't need pre generated HD maps is because their system generates said HD maps on the fly.

Why waste resources when the car can do it in realtime and better accuracy then pre-loaded HD maps. Nothing can beat realtime lol.

I just got FSD beta, and while the driving "Logic" is not there yet, they definitely have the vision system down. Its insane how good the vision system is at building the world it sees around it and how accurate it is.

 

[S4WRXTTCS]

if what i said was so garbage then Tesla would have correct maps already and not have parts of its maps outdated by months/years.
Unfortunately crowdsourced detailed map that is automatically updated is the only way to have correct maps at scale

I think the entire industry is going about Map data incorrectly.

The only real solution I see is for the federal government to standardize on a Detailed Map database that's in the public domain where every road construction company was required to update the maps any time they changed anything.

Where that Map database itself acted to authorized autonomous driving over that road. Something so detailed that in a sense it's a second earth. Crowd sourcing changes is a necessary part of it as there will always be things like pot holes or things damaged by crashes.

Relying on crowdsourced updates will always result in incidences as things won't get updated until something has already transpired. Like with Apple Maps I report stores that are closed due to a store closure or due to a change in hours, but I'm already at the store when I do it. Unless I was smart enough to check google maps before I left.

That action itself means I'm checking two databases instead of one.

Mapping is something that a single company can't just solve. It requires the federal government, and the private sector to work together.

 

[EVNow]

The only real solution I see is for the federal government to standardize on a Detailed Map database that's in the public domain where every road construction company was required to update the maps any time they changed anything.

LOL - as impractical as any fairy tales HD Map apologists spin here. I can see how quickly some TV channels can turn it into a political battle.

The real long term solution is to have robust maps but for the car to be able to drive without the map being up to date for short distances (with remote or in car confirmation).

 

[RDoc]

I think the entire industry is going about Map data incorrectly.

The only real solution I see is for the federal government to standardize on a Detailed Map database that's in the public domain where every road construction company was required to update the maps any time they changed anything.

Where that Map database itself acted to authorized autonomous driving over that road. Something so detailed that in a sense it's a second earth. Crowd sourcing changes is a necessary part of it as there will always be things like pot holes or things damaged by crashes.

Relying on crowdsourced updates will always result in incidences as things won't get updated until something has already transpired. Like with Apple Maps I report stores that are closed due to a store closure or due to a change in hours, but I'm already at the store when I do it. Unless I was smart enough to check google maps before I left.

That action itself means I'm checking two databases instead of one.

Mapping is something that a single company can't just solve. It requires the federal government, and the private sector to work together.

Well, based on other government efforts I'm pretty doubtful such a system would be real time for any reasonable definition. Why would some bureaucrat bother to do it today when next month is good enough?

 

[S4WRXTTCS]

Well, based on other government efforts I'm pretty doubtful such a system would be real time for any reasonable definition. Why would some bureaucrat bother to do it today when next month is good enough?

Road Construction companies would be required by the ruling to update the publicly accessible database as the construction took place.

There are all kinds of regulations set forth by governments around the world where private companies have to follow or face fines.

This is also something that would help construction companies as it would give them the ability to send an update to one database that companies like Google Maps could pull from.

 

[EVNow]

There are all kinds of regulations set forth by governments around the world where private companies have to follow or face fines.

Nothing wrong with the idea except it’s difficult to do at federal level because of partisan politics.

BTW, map makers already use local government sources. This is from TomTom.

For example, we work with local governments and road authorities to understand planned changes and use our global community of 600 million connected devices to detect changes that were perhaps unplanned. These changes are fused to ensure the most accurate reflection of reality.​

 

[Supcom]

The only real solution I see is for the federal government to standardize on a Detailed Map database that's in the public domain where every road construction company was required to update the maps any time they changed anything.

Sort of like this database?

Not that construction companies update Tiger, but would you really want them to? It would seem more like a job for the city/county/state road departments. After all, the construction of the road is only one part of this. You need someone to update all the traffic lights. stop signs, school zones, No right turn on red intersections, no passing zones, etc, etc, etc. And you need it done in a timely manner, a characteristic universally shared by governments worldwide.

Oh yeah. This has to be done globally.

 

[EVNow]

And you need it done in a timely manner, a characteristic universally shared by governments worldwide.

LOL.

BTW, I chuckle when someone says how inefficient government is compared to private companies. I have first hand knowledge of how inefficient private companies can be …

 

[Supcom]

LOL.

BTW, I chuckle when someone says how inefficient government is compared to private companies. I have first hand knowledge of how inefficient private companies can be …

I'll bet that the Dept of Commerce contracts out Tiger updates to private companies.

 

[S4WRXTTCS]

Sort of like this database?

Not that construction companies update Tiger, but would you really want them to? It would seem more like a job for the city/county/state road departments. After all, the construction of the road is only one part of this. You need someone to update all the traffic lights. stop signs, school zones, No right turn on red intersections, no passing zones, etc, etc, etc. And you need it done in a timely manner, a characteristic universally shared by governments worldwide.

Oh yeah. This has to be done globally.

I can't see it happening globally, but I do think it would be an important step in trying to make sure China doesn't overtake us in the race to autonomous cars.

Construction companies need access because they have to close roads, and so they should be able to update the database with what roads are closed and what time they are closed. It wouldn't be just used for autonomous cars, but it could be something google or apple maps pulls from.

Now obviously things like actually authorizing autonomous cars for a road has to be handled by the city/country/state along with making any changes that they make. Like near where I live the city changed the rules of the middle lane, but Apple maps still has the old rules.

This type of singular database would also help counties maintain their roads.

For a long time I assumed where I live (in Snohomish county) would be slow at fixing pot holes, but it actually tuned out that they were really quick. So quick the pot holes were fixed within a couple days of reporting them.

I shouldn't need to report it as it should be my car that marks it, and reports it.

 

[TheTheocracy]

I wonder if the problems are not specifically a data problem but a training/method problem. Given that Tesla's self-driving stack is mainly deep NNs, it is hard to figure out what truly is the problem in misbehavior cases. Either it's empirically shown through percent accuracy or assumed through specific cases. Since I guess they are naturally data-driven with all of the algorithms, training in new data is more empirical and may or may not mitigate the problem.

Which part of the stack is causing the problem? What training data should be used to retrain the models (thinking of class imbalance and training bias)? Is the validation set too small or not comprehensive enough to encompass mitigation of local problems? Does retraining part of the stack affect other parts of the stack? I'm not so sure if the problem and solution is clear cut.

A problem with solving more localized problems is the creation of local NN models, which requires loading in different models when traveling to different places. Maybe there is a way to store a generalized model and that same model retrained for your local area?

Maps with local problem data does not account for changing models. Retraining NN models may change where weird behavior occurs. In many cases, they start with a new model and retrain with the new data. Are the maps used for more pattern of life analysis for going through intersections and the like? How about changing backgrounds, occlusions from new structures, etc.? The merger between algorithms and data is not as trivial as we'd like it to be.

Hopefully we can find some sort of solution that shows both algorithm performance, mitigation of key problems, and why it's working to inform future algorithms and data training. Maybe I'm wrong and more data is the way to go. That would be great! I'm just speculating with all of these things anyway.

 

[EVNow]

@TheTheocracy

Depends on the problem - but a lot of the issues are planner issues rather than NN. There are probably some issues wiith prediction as well - which I guess is partly in NN and partly in the planner.

But I do hope after all this time devs can easily triage and figure out the cause of any issue we report easily. But solutions may not be simple …

 

[TheTheocracy]

@TheTheocracy

Depends on the problem - but a lot of the issues are planner issues rather than NN. There are probably some issues wiith prediction as well - which I guess is partly in NN and partly in the planner.

But I do hope after all this time devs can easily triage and figure out the cause of any issue we report easily. But solutions may not be simple …

Yeah that makes sense. I've seen the push for NN-based planner, which would be cool. The current solution of a coarse search and a continuous optimization function, depending on the situation, seems like it may oscillate given a set of parameters and when situations change. I wonder if it is blessing for our reaction time to be slow in some cases because we don't react impulsively based on possibly faulty perceptions. It seems like the planner based on other agents and their actions doesn't have the correct weighting or training set. And the reactivity to other agents is good in some cases and terrible in others. I wondering if planning isn't good because of other agents, then the extraction of information on the other agents doesn't work as well or the assumptions created about the agents are not correct. But that would have to be a deeper analysis by the AI team.

 

[kabin]

Any ideas how Tesla is using NNs? I can't imagine endlessly training a NN (or multiple NNs) to appropriately respond to millions of roadway scenarios. The NN supercomputer training program might never get a break. NNs work well for pattern recognition so I would guess they would process video data to identify and give locations and possibly velocity/acceleration for vehicles, people, bikes, cross walks, roadways, roadway signs, orange cones, traffic lights, debris, etc. and higher order logic or another NN would classify the scenario and to then appropriately control the vehicle. That might be like herding cats.

 

[linux-works]

Do you prefer occasional phantom braking vs occasional crash ?

Basically false positive vs false negative.

we have 'occasional crash' right now with human drivers but almost never, other than learners, do PB.

autonomous or even semi, should not be released until the PB problem can be managed, somehow. perhaps that's a learned stretch of road where the user says 'no! dont ever do that here'. right now, tesla does not learn from user driving, not like that. that would be at least one bandaid.

but really, if the feature is not ready, its not ready. and really - for most of the world, its not ready. I live in the bay area where its over-trained so I benefit from that. tbh, its one reason I bought mine; I knew I'd have a better driving experience than most other geo's simply due to tesla density here.

but I get the fact that PB happens a lot outside the bay area. that, to me, means, the tech is far from ready for widespread deployment.

I fear the day when teslas can be rented and cost similarly as ice rentals. regular people (lol) are just not ready for this. enthusiasts just BARELY are.

 

[TheTheocracy]

Any ideas how Tesla is using NNs? I can't imagine endlessly training a NN (or multiple NNs) to appropriately respond to millions of roadway scenarios. The NN supercomputer training program might never get a break. NNs work well for pattern recognition so I would guess they would process video data to identify and give locations and possibly velocity/acceleration for vehicles, people, bikes, cross walks, roadways, roadway signs, orange cones, traffic lights, debris, etc. and higher order logic or another NN would classify the scenario and to then appropriately control the vehicle. That might be like herding cats.

I think it is like how you said. Transferring all image features using NNs to a vector space as objects/trajectories and use a logic or another NN for planning the route. My guess is they are using some sort of reinforcement learning approach or genetic algorithm for the eventual NN planner. I can't imagine putting all of the constraints, rewards/penalties, etc. into training the system as it is tough to create a robust environment to train in. Works for games so it should work for driving?

but really, if the feature is not ready, its not ready. and really - for most of the world, its not ready. I live in the bay area where its over-trained so I benefit from that. tbh, its one reason I bought mine; I knew I'd have a better driving experience than most other geo's simply due to tesla density here.

but I get the fact that PB happens a lot outside the bay area. that, to me, means, the tech is far from ready for widespread deployment.

There is a lot of research in human machine teaming that suggest trust in an autonomous solution is very important in the usage of a solution. Some over-trust a system or maybe the system behaves in unexpected ways as to undermine the trust of the system. Managing an appropriate level of trust or even the system communicating an appropriate level of trust is really helpful in all of this (we need explainable AI). Maybe this can be something that Tesla can address? Or maybe there can be a simple front TACC instead of normal TACC?

It would be amazing to see all of the detections and paths during a PB event to really see what goes on. I don't think that we can deny that the system works, but those PB events really do make using it less enjoyable.

 

[linux-works]

I can't remember which person was talking about this: could have been dan dennett or robert sapolsky or maybe jaron lanier (all authors I follow a lot, lately on YT) - but they were talking about doing translation and how machine 'learning' really needed to be able to re-sample the knowledge base pretty much every day, since new phrases or pop culture memes pop into existence and any translator would have to re-import its NN logic, essentially, constantly. ie, it never completes and is always out of date.

the more I hear scholars talk about 'AI' the more I'm convinced that this isn't how we are going to get self-driving cars to be safe enough to deploy in the chaotic world.

AI is a party trick and it gets us to a level we didn't easily attain before (in translation, for one example). big data was 'needed' but it still runs into walls and wisdom seems to indicate that its still yet another intermediate step in how we try to understand thought and mindful computation.

the AI data model is sloppy, too expensive and its not going to get us there. that's my xtal ball reading, at least.

 

[linux-works]

It would be amazing to see all of the detections and paths during a PB event to really see what goes on. I don't think that we can deny that the system works, but those PB events really do make using it less enjoyable.

visions of feynman diagrams and removing infinities...

that's it! we have to have the computer solve all possible paths at once, and then the solution becomes obvious.

how many qubits does the model 3 have, again? I forget.

 

[TheTheocracy]

the more I hear scholars talk about 'AI' the more I'm convinced that this isn't how we are going to get self-driving cars to be safe enough to deploy in the chaotic world.

AI is a party trick and it gets us to a level we didn't easily attain before (in translation, for one example). big data was 'needed' but it still runs into walls and wisdom seems to indicate that its still yet another intermediate step in how we try to understand thought and mindful computation.

the AI data model is sloppy, too expensive and its not going to get us there. that's my xtal ball reading, at least.

I don't blame you with thinking current AI is a party trick. Current deep learning techniques are more or less similar to previous years techniques and hasn't really changed the paradigm of learning. I agree with you that the AI data model really is sloppy. I just hope you don't lose the wonder of trying to finding a solution.

visions of feynman diagrams and removing infinities...

that's it! we have to have the computer solve all possible paths at once, and then the solution becomes obvious.

how many qubits does the model 3 have, again? I forget.

Haha, maybe I wasn't clear enough? I was referencing how the current planner in FSD finds optimal paths. Maybe we can see what those look like during FB?