-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
The Tesla / Mobileye story
- Thread starter lunitiks
- Start date
Sunshine State
Automotive Enthusiast
Still such a shame though that the customers have to suffer when Tesla cars could clearly have been so much further ahead with Mobileye. The really strange part is that whenever Tesla releases information or more acutely "spins information" everyone has to try to figure out what the truth actually is.
Last edited:
I have information from an insider inside the admin..... (TOP SECRET) Trump loves getting peed on.This is a wonderful and interesting history but I’m too busy reading this
Does anyone know if Hotz ever softened his view that Mobileye is a failing company? Part of me wonders if Hotz blamed the folks at Mobileye for planting enough seeds of doubt with Musk to sabotage working with Hotz. That story is important to since Hotz has insinuated that Musk liked what he had to say, and indirectly takes credit for his interactions with Musk playing a big role part in the Musk's decision to part ways with Mobileye.
Don't know about that, but I do agree Musk's public commentary and Hotz seem to share the same very public belief about autonomous being solved through aggressive deep learning, instead of the more nuanced approach advocated by the likes of MobilEye, Waymo or Audi, for example, that has included building of rules and human trained elements.
In the end I believe all of them are probably converging behind the scenes much more than people think. Nobody is likely to do a black box all-seeing-all-driving NN, even the newer players are likely to have rules and algorithms in their systems, and on the other hand the older, more rules-based, human-trained players in the field are moving more and more into deep learning / NNs.
In the end I believe all of them are probably converging behind the scenes much more than people think. Nobody is likely to do a black box all-seeing-all-driving NN, even the newer players are likely to have rules and algorithms in their systems, and on the other hand the older, more rules-based, human-trained players in the field are moving more and more into deep learning / NNs.
I’m in total agreement. I have AP1 and AP2 in the garage. IMHO, AP2 is better now than AP1... and getting better. Don’t get me wrong AP1 is still great.
PS would love to hear your experiences with AP2-3 on Model 3. Probably better in another thread:
.48 feels like AP2 finally passed AP1
... mjeee... I don’t know. My impression after putting together the bits and pieces in my timeline, was that some awkwardness between the two companies started fall 2015. Tesla is obviously gearing up for HW2.0 spring 2016, hiring a bunch of fancy schmancy people while preparing for M3 unveil.
That reported visit Musk did with Mobileye in March could’ve been when the line was cut.
Just speculating of course. I just don’t see that Tesla was taken by any great surprise.
Btw I also disagree that Tesla was no important customer to Mobileye. IMO mobileye got a huge boost from Tesla. Personally I had never heard of the company until HW1.0 came out.
Not disputing Mobileye has extremely important customers in other brands of course
I just said the theory is plausible, not necessarily my theory or only plausible theory or even the best theory. As said, I think both sides negotiated hard for something they wanted and failed to get it. I also think both sides probably were surprised at some point in the process.
This probably played some part in Tesla skipping whatever they had planned for Model X launch or Model S facelift at the latest (additional cameras/radars). March 2016 was around Model S facelift unveil time, so they probably knew already earlier that was not happening (possibly a disappointment for MobilEye) and were focused on whatever was the next step - probably Tesla negotiating EyeQ3 for their AP2 board, which resulted in a disappointment for Tesla...
My theory:
- For whatever reason Tesla didn't do or want to do the "AP 1.5", which was planned for Model X launch (September 2015) and then for Model S facelift launch (April 2016). Maybe it was delayed, maybe they wanted that EyeQ3 raw access, etc. This was probably a disappointment for MobilEye. This would have had at least the dual front cameras (up to three cams) and rear corner radars.
- In the meanwhile Tesla had been working on the next level, AP2, with Nvidia. DRIVE PX 2 schedule means this could happen late 2016 at the earliest, but late 2016 was fast approaching as "AP 1.5" had been delayed. This AP2 plan also included an EyeQ3 chip (similar to Audi zFAS combining Nvidia DRIVE PX/MobilEye EyeQ3) at least on the first year or few of the board. For this, MobilEye probably required some access Tesla was unwilling to give on the driving data side.
- Tesla is also working on its own silicon for eventual AP2 successor.
- Disappointments abound on both sides come late spring 2016, for both "AP 1.5" cancellations and for the negotiations being tough on EyeQ3 on AP2. There is still talk probably of EyeQ3 on the AP2 board, because Tesla would need it for the transition, and well, MobilEye would probably prefer to sell chips rather than not. Negotiations are not going well though, then Joshua Brown happens. MobilEye decides this is not worth it, Tesla decides AP2 is close enough and they can spin the rest even without Q3, since AP2 was already going to ditch it eventually (just not yet).
- AP2 and the EAP debacle happens, because Tesla has to emulate EyeQ3 in their code (which they do, in a limited fashion), and leave an empty slot on the board where it likely was instead of using EyeQ3 to drive EAP initially.
That's my speculation/theory.
But everyone in the automotive industry, their actual customer base, had. I was driving a MobilEye car already back in 2010 (and I didn't even know it then).
I think it is too much of a reach to postulate on how things are converging without specific background in deep learning, @jimmy_d has been invaluable on providing real context and insight. However, I love this thread because of the psychology of it all. Reminds me of the double helix book by Watson talking about the race among scientists and how these big leaps look behind the scenes. Elon is keeping his hands close the vest on Autopilot progress and this thread has made me think more and more this is for a good reason. In my opinion, Tesla needs to keep leading in this area with trying to bring a parsimonious product with self driving integrated and usable by even my tech averse elderly aunt who somehow works with an iPhone now when she still can’t power on a dvd player
@AnxietyRanger You'll like this idea (I think). Similar to @Cosmacelf .
Unfounded opinion of mine: Tesla was going to use Mobileye both to have Autopilot today (similar to the rain sensor discussion, but with opposite approach) and to train Tesla's AP's NN. Mobileye freaked out since this would end up making them obsolete, so they used the Joshua event as a reason to part ways.
Unfounded opinion of mine: Tesla was going to use Mobileye both to have Autopilot today (similar to the rain sensor discussion, but with opposite approach) and to train Tesla's AP's NN. Mobileye freaked out since this would end up making them obsolete, so they used the Joshua event as a reason to part ways.
Let’s not neglect the MASSIVE elephant in the room as well.... how much did Mobileye make on the deal as they parted with Tesla?
Many factors yes, but never underestimate the power of a crap ton of money and a promise of more directional control of how your self driving approach should progress
Many factors yes, but never underestimate the power of a crap ton of money and a promise of more directional control of how your self driving approach should progress
Tesla seemed to think they could do FSD very quickly with just cameras. They were wrong, the 2017 demo didn't happen and as the lack of motorcycle detection in AP2 shows it's not that easy to replicate what MobileEye did with radar.
Tesla seems to be banking on FSD being a bunch of simpler technologies. Follow roads and make some lane changes, use auto-park, use GPS navigation and you are there, right?
Everyone else, MobileEye included, doesn't seem to think it's that easy. That's probably why they split, Tesla wanted it done by 2018 and MobileEye had a more realistic timeframe in mind, with the knowledge that any affordable hardware today wouldn't ever do FSD.
Tesla seems to be banking on FSD being a bunch of simpler technologies. Follow roads and make some lane changes, use auto-park, use GPS navigation and you are there, right?
Everyone else, MobileEye included, doesn't seem to think it's that easy. That's probably why they split, Tesla wanted it done by 2018 and MobileEye had a more realistic timeframe in mind, with the knowledge that any affordable hardware today wouldn't ever do FSD.
Possible, though not the likeliest reason in my mind. I mean, nothing would stop Tesla from plugging an EyeQ3 in an AP2 NN training board (maybe they already do)... But sure, maybe that would have been even easier with a raw stream.
stopcrazypp
Well-Known Member
I think the difference is Audi wasn't trying to duplicate any of the Mobileye chip functionality in-house. The Drive PX hardware is for sensor fusion, but not for replacing EyeQ.
Even though Tesla is also using the PX hardware, that's not really the significant issue Mobileye had. The main issue Mobileye had was Tesla was developing the Tesla Vision software (which Elon alluded to being flexible enough to be used with different hardware, not just Nvidia's), which is a direct replacement for the EyeQ3 chip.
Possible, though not the likeliest reason in my mind. I mean, nothing would stop Tesla from plugging an EyeQ3 in an AP2 NN training board (maybe they already do)... But sure, maybe that would have been even easier with a raw stream.
On a couple cars sure, but they could no longer bulk purchase the chips/ sensors.
Well, they can obviously to maintain the AP1 fleet... and how many would they need? 10-20 would be plenty for NN training...
Possible. Though I'd clarify I believe on zFAS the driving at speed sensor-fusion is actually done outside of DRIVE PX anyway. I believe DRIVE PX does the parking, EyeQ3 does the forward driving seeing and then a third chip does things like Lidar fusion.
Similar threads
