Will Lucid license Tesla FSD?

[lucidukan]

I paid for the DreamDrive package and so far the only feature ie Lane Keeping with a rudimentary Lane change. There does not seem to be anything in the horizon. No ETA for anything close to the Tesla FSD. Soon my hardware will likely be obsolete.

Any chance Lucid will license the tech from Tesla?
I hear some rumors of OEM licensing deals and Musk has floated the possibility in the past. This might be the fastest route for Lucid even if it’s temporary till they build their own stack.

 

[HC_79]

Tesla has bet everything on FSD so I don’t see any deals happening for the foreseeable future unless someone is prepared to pay some serious $$$$. They give it away they’ve lost the only advantage they have over the competition because they’re certainly not innovating on the EV front anymore.

Lucid just partnered with some company in Saudi Arabia to make use of AI and gain access a super computer. They’ve admitted ADAS is a weak point and with this deal they’ll be able to speed up the development apparently.

 

[dwfallin]

They also just announced hands free for dd pro - isn’t that ‘something on the horizon’?

 

[boydracing]

Tesla has bet everything on FSD so I don’t see any deals happening for the foreseeable future unless someone is prepared to pay some serious $$$$. They give it away they’ve lost the only advantage they have over the competition because they’re certainly not innovating on the EV front anymore.

Lucid just partnered with some company in Saudi Arabia to make use of AI and gain access a super computer. They’ve admitted ADAS is a weak point and with this deal they’ll be able to speed up the development apparently.

I agree with HC_79 there is no way for Lucid to partner/license FSD from Tesla. There are too many technical leaps going on right now within AI (software) and very soon quantum (hardware). We will see AI get way better by leveraging quantum in the future. The data that is fed into AI will only get more accurate.

 

[Ktw13]

FSD, though with fancy name and social advertisement is lvl2 at best, no one knows if vision only is a dead end pr not.
Many car companies use Nvidia platform to develop their ADAS as there is minimal trade secrete exchange.
With Tesla, Lucid or any other car company with valuable technology would not make themselves vulnerable involving competitors software in their system.

 

[WhistleKing30]

I don't see Lucid adopting FSD any time soon. Definitely not for their current line-up (Air and Gravity) as this would require to change the entire hardware stack to Tesla's. Which is mighty expensive. Tesla's current algorithms rely heavily on a very specific set of hardware. Too many changes would require re-training the models, which kind of defeats the purpose. Assuming the current fleet is capable of uploading traffic snapshots, it would probably be cheaper for Lucid to hire the necessary server capacity from Amazon and train their own neural nets for the selected regions. It's not rocket science.

An alternative is to pay Tesla to re-train the models based on Lucid's current sensor suite. Though I believe will be an unlikely scenario, unless time-to-market is the only criteria that matters. Which it might be. But I seriously doubt it.

If there ever would be an FSD licensing deal then this would more likely be for a new platform. For example Lucid's upcoming mid-sized car. Though if Lucid chooses the path of purchasing a license instead of developing their own then the more likely candidates are the Tier1 suppliers such as MobileEye. I believe.

 

[JMason]

One fly in the ointment I see is Musk has considered Lidar anathema so I doubt there is any code in Tesla FSD for radar input. While Lucid’s DDPro/Premium seems clunky in some comparisons to Tesla (of which I know nothing having never driven a Tesla) a Lucid will not drive through a painted wall (infamous Road Runner Test) as Tesla’s FSD does. Cameras alone are inadequate and, given that, radar is a huge advantage over Tesla’s implementation of FSD for me. From what I’ve read the primary reason Musk is so anti-lidar is cost.

Access to a supercomputer and AI hopefully will address shortcomings in DDP quickly. Having Hands Free Driving announced for the end of this month (July) is a plus. Still, not sure I’ll be turning loose of the steering wheel though

 

[illopp00]

Well they just announced a licensing deal with Nuro.

 

[Steveinarizona]

It wouldn't make any sense to me. Leaving aside FSD's safety issues, it is based on camera use only. Lucid's system is predicated on cameras, sensors and lidar. The conversion costs are probably as great as the creation cost for Lucid.

VW invested $5 billion in Rivian for its use of and conversion of the Rivian software for future VW vehicles. Given VW's history, I wouldn't give much chance of it being a successful effort.

AI is moving so quickly that I suspect that it will allow for much accelerated improvements to ADAS systems. To me, creating in today's world is a better use of funds than converting.

 

[WhistleKing30]

While Lucid’s DDPro/Premium seems clunky in some comparisons to Tesla (of which I know nothing having never driven a Tesla) a Lucid will not drive through a painted wall (infamous Road Runner Test) as Tesla’s FSD does. Cameras alone are inadequate and, given that, radar is a huge advantage over Tesla’s implementation of FSD for me. From what I’ve read the primary reason Musk is so anti-lidar is cost.

Automotive radar would not recognize the wall in the road-runner test either. Though LiDAR would.

The question is, however, what does failing the road-runner test really tell us? I mean, how often do you (as a human) encounter a painted wall on the road? Is this a real problem? And secondly, what would we humans do when driving towards a tunnel entrance which turns out to be a super-hyper-realistic painting on a solid wall? My bet is that quite many would speed towards it before it is too late.

I know it's an unpopular opinion but Musk is not wrong about visible-light cameras being sufficient for human-like driving. Though an important condition is that the algorithm is also human-like smart to interpret known and unknown environments based on this visual data. In Elon-years this will happen some time in 1974 2026. But realistically, I think that it will be closer to a decade before we see true level 5 autonomous driving using only cameras. Until then, I believe that the use of LiDAR and other sensors is necessary. In that sense Lucid is in a better position.

 

[Leftlane]

Don’t underestimate Elon’s pettiness either - I think he would gladly pass on the $$$ to not let Lucid get FSD from Tesla.

 

[knewman]

The question is, however, what does failing the road-runner test really tell us? I mean, how often do you (as a human) encounter a painted wall on the road? Is this a real problem? And secondly, what would we humans do when driving towards a tunnel entrance which turns out to be a super-hyper-realistic painting on a solid wall? My bet is that quite many would speed towards it before it is too late.

You're right, it's an abstract scenario. But the intention of the scenario is to highlight the lack of data the system has to work with before we even get to any sort of logic. The classic "garbage in, garbage out" still applies in today's computing, and the consequences are even more extreme at 70mph.

I know it's an unpopular opinion but Musk is not wrong about visible-light cameras being sufficient for human-like driving.

Again, you're dead-on that visible-light cameras will probably match average human behavior in typical situations. The problem is, software cannot match the decision-making power of the brain. Not even close. The strength of automated systems is the speed at which choices are made based on the available information, with reliable, measurable, reproducible output. Computers are fantastic at that. But the decisions lack accuracy, principally, among other points, because software is fundamentally unable to make inferences about what's happening around it. Because of this, automated driving systems must far surpass human ability before they can be a realistic concept at scale. Cameras aren't sufficient to achieve that goal.

 

[segbrk]

Well they just announced a licensing deal with Nuro.

Except there has been no such announcement? The only announcement is that Uber is buying a bunch of cars, presumably with some Lucid support to integrate Nuro's software and hardware into Uber's vehicles. It doesn't mean any of Nuro's software is coming to the rest of us. Nuro's software is presumably bound to Nuro's funky sensor hat and wouldn't be trivially applicable to Lucid's sensors in normal consumer cars.

 

[WhistleKing30]

You're right, it's an abstract scenario. But the intention of the scenario is to highlight the lack of data the system has to work with before we even get to any sort of logic. The classic "garbage in, garbage out" still applies in today's computing, and the consequences are even more extreme at 70mph.

My objection with the roadrunner experiment is that reality doesn't really provide garbage-in scenarios. Thereby making this experiment rather pointless. More specifically, why should a system need to deal with scenarios that are not in any way a proxy to reality? Wouldn't it make much more sense to subject FSD to a more realistic edge case? Just an example I can think of right now; placing a window with reflective coating at a 45 degree angle to a tangent road - thereby giving the illusion of an unobstructed road inside the window. That would - arguably - be a much better proxy to the system performance in a roadrunner-like scenario.

Neural nets such as FSD are trained on trillions or data points. If none of these trillion data points contained a given scenario (ex. roadrunner) then the algorithm will simply not have any reference to work with. But the bigger the training dataset, the less likely it becomes that an edge case is not covered. That's why the real complexity in achieving NN-based autonomous driving is not the theoretical complexity (because actually it's quite simple) but rather the sheer size of the dataset that needs to be acquired and processed correctly. Tesla has previously invested heavily in fleet-capable data collection (e.g. each Tesla on the road can upload snapshots of traffic scenarios) combined with purchasing in-house computing power (Tesla Dojo). Today they have a competitive advantage. Exactly because of the model size that I mentioned earlier. And the number of edge cases is only getting smaller by the day.

If Lucid would join the autonomy race today, they would be late to the party, yes. But at the same time Lucid would also have a massive advantage because a large portion of the fleet is equipped with LiDAR. I believe that these additional data points are so valuable that it may just allow Lucid to leap-frog into a viable AD solution on par with FSD at a fraction of the price. Even with the huge difference is fleet size when compared to Tesla, just because the quality of the data in so much higher in Lucid's current fleet.

Again, you're dead-on that visible-light cameras will probably match average human behavior in typical situations. The problem is, software cannot match the decision-making power of the brain. Not even close. The strength of automated systems is the speed at which choices are made based on the available information, with reliable, measurable, reproducible output. Computers are fantastic at that. But the decisions lack accuracy, principally, among other points, because software is fundamentally unable to make inferences about what's happening around it. Because of this, automated driving systems must far surpass human ability before they can be a realistic concept at scale. Cameras aren't sufficient to achieve that goal.

Not sure I agree. There will be a performance gap for the foreseeable future, no doubt. But should that stop us from accelerating technology that will eventually lead to zero-casualty traffic, while at the same time giving our economy a tremendous boost? I'm not giving an opinion here. Just pointing out that this is a philosophical discussion where it's up to our legislators to determine what is safe enough. And what's safe enough is not necessarily "safer than human" in every category.

 

[tarator]

It seems that some people think that access to supercomputers is the bottleneck for making good autonomous driving software. Let me just remind them that people that know how to do that are not that many and are worth few hundred millions a pop.

Lucid doesn't have the team to do anything in this area. They can't afford to.

 

[WhistleKing30]

It seems that some people think that access to supercomputers is the bottleneck for making good autonomous driving software. Let me just remind them that people that know how to do that are not that many and are worth few hundred millions a pop.

Where do I sign up??
All kidding aside, yes, practical implementation is not trivial. But for us who have worked in the autonomous driving space will recognize that the biggest challenge really is finding the balance between model complexity and performance. It's not a one-man job, but rather three-team-job within the holy trinity: classification, planning and actuation.

There are a number of self-driving startups out there that are doing marvelous things without having a multi-billion-dollar budget. Though I will admit that it's the last 10% of race that takes 90% of the efforts (and talent). And that kind of capex is currently reserved only to the industrial giants. But the space is moving at lightning speed, so who knows what the upcoming 2-3 years will bring.