Tesla is not running in the same race as actual AV companies.
Operators like Waymo do look plausible while operating within their tiny geofence. Driving pre-validated roads does impress some people. Right now if you dropped a Waymo and Tesla FSD on an unsealed road in Michigan, one of them will drive at least as well as a human learner driver and the other will probably refuse to move.
The point is, I agree that there could be many firms which are far ahead of Tesla, but there isn't enough information in the public domain to say this with any confidence. Right now Tesla's is the only major system which can be assessed with critical objectively.
If it was just one single trip, I was incapacitated and just needed to get home, then fine, level of human learner driver is good enough. But for everyday usage? No way, the risk is just too high.
refusing to move is a feature not a bug. and human learners don't slam into a jersey barrier because there's some glare.
why tesla bros continue to insist that they have a good vision stack while they can't even do contrast adaptation is beyond me. the dril tweet comes to mind
Also your claim about contrast adaptation seems awfully specific, enough that I'm interested in a citation. There's plenty of evidence that their imaging sensors do adapt well in high dynamic range circumstances.
They literally do, which is why jersey barriers are literally designed with vehicular impact in mind.
> contrast adaptation of CCDs
You can't just mix terminology like that. Contrast adaptation describes behaviour in biological systems which isn't analogous to anything a CCD does. There's a reason why literally nobody uses the term to describe digital systems, because it's not relevant.
If you were trying to say that the CCD doesn't have sufficient dynamic range, then just say that instead of using a term which doesn't describe dynamic range.
One will drive and likely get into an accident, and the other will responsibly not do what it wasn't designed for. But if you took away Waymo's responsible disposition, I'm sure it would drive better than the Tesla, because it's just way further along in development, more and better sensors, more rigorous testing and simulations, more corner cases that absolutely had to be dealt with to roll out a real self-driving product.
What are you basing this prediction on? There is extensive independent documented video evidence of FSD beta driving and navigating on unsealed roads in Michigan with no human intervention required. This includes driving at night, in the rain, and with a diverse array of obstacles.
> [Waymo is] just way further along in development
This may well be true. However there's no publicly available information to make any objective assessment. For better or worse (and I see valid arguments both ways) Tesla is airing their dirty laundry in public for all to see. It can be assessed by objective third parties in privately owned vehicles without the blessing or oversight of the parent company.
By comparison we have no way of assessing Waymo's stack objectively. For example we don't know to what extent corner cases are embedded into the software or handled by humans remotely. Was the 15th Avenue incident resolved by human intervention or an improvement to the stack? To what extent is their vision and planning stacks capable of universality or over-fitted to regional specifics? No objective analysis is possible, so no comparison is possible.
What laypeople don't seem to get is that AV mistakes don't follow the same "scale of alarm" as human mistakes.
Tesla is making mistakes that, for a human, might not be world ending, but for an AV make absolutely no sense.
Things like recognizing it needs to make a turn and making no attempt at all to slow down: https://www.youtube.com/watch?v=6wGAyAN3iRo
Or recognizing a vehicle is in its path and not even trying to slow down: https://youtu.be/6wGAyAN3iRo?t=21
Or recognizing a stop sign and ignoring it to follow the car in front: https://youtu.be/21ShKWLAnWA?t=548
Classification issues in AVs are supposed to be the achilles heel, but Tesla is still running into situations where perception works correctly and the AV just completely ignores ground rules that any serious AV must be built around.
> Was the 15th Avenue incident resolved by human intervention or an improvement to the stack? To what extent is their vision and planning stacks capable of universality or over-fitted to regional specifics? No objective analysis is possible, so no comparison is possible.
It doesn't matter when the product you're comparing it to ignores stop signs. Tesla's FSD approach is the definition of the local maxima problem for AVs. It makes progress on the axis that impresses lay people by sacrificing progress on the axes that matter to the long term success of an AV. Geofencing and arbitrary sensor limitations are just the tip of that iceberg...
That's because they aren't being reckless and releasing a dangerous product onto the road that they know has issues. Unlike Tesla.
And those geofenced areas are at least with Cruise almost entire cities.
Until they fix that, there will be lots of limitations. It could get to be good enough to be hands off a lot more than it is now, though.
I'd like to see what happens if someone did the following to test the assertion that "cameras are all we need". Cover all the windows of a Model S with LED screens, and route all the cameras to display onto them. Then see how well human drivers do with that visual input which roughly simulates the fraction of a second delay to perform the FSD processing, and removes our stereoscopic perception when we turn / move our heads around.
Many software and hardware engineers do not appreciate the amount of work done by evolution upon our kinesthetic / proprioceptive packages. Sure there is wide variability in its implementation (that friend you won't under any circumstances let behind the wheel with you in the vehicle). But by and large, it performs under a set of energy and materials constraints that continues to school our current engineering state of the art.
So I sometimes also wonder if the ultrasonic/sound sensors in Teslas are sufficient to reproduce the control system our bodies' proprioception engages when we drive.
Why can that never happen?