I think we’re on the same page here. Robotaxi will require human supervision throughout its ODD until it’s not shitty anymore, which for L4 is an extremely high bar. Mechanical failures (e.g. flat tire) should vastly outnumber autonomy failures in practice before Robotaxi could be considered remotely [see what I did there] ready for driverless operation.
FSD should have observed the lead car swerving, realized that it probably wouldn’t be doing that for no reason, and defaulted to a similar lane repositioning, even before the obstacle became visible. In this case only a very minor repositioning would have been needed to avoid the bucket. The E2E highway stack will very likely solve most situations like this, or at least I hope it will.
It’s a moot point until Tesla has logged millions of miles with pure vision L4-in-training. At that point, the nature and rate of its failures can be compared with the nature of L4-with-Lidar failures from other manufacturers, and patterns may emerge. If the patterns are strong enough, and if a strong technical case can be made that Lidar would have prevented a significant number of Tesla’s observed failures, then that could lead to a Lidar mandate. We’re in early days yet. (Prediction is difficult, especially when it’s about the future!)
The same way that a human would. (Drawing on intuition based on years of driving experience.) This is the strength of E2E deep networks; that they simulate such intuitive processes.
FSD will be trained over millions of examples to perform the driving task as safely and effectively as possible. It does not require an explicit notion of “when to follow the lead car” or “is that driver falling asleep”, just as our instinctive driving reactions (after years of experience) don’t need to consciously follow such logic either.
The irony with this particular failure is that in the early days, Elon used to crow about how Autopilot could effectively see past the lead car, by bouncing radar signals underneath. Sigh.
I have been in such an accident in my early days. I was riding shotgun. Too late to make any decisions and we ended crashing into the object that the driver ahead of us swerved at the last minute to avoid.
Not every problem is a trolley problem. (The vast majority are not.) FSD has constant 360-degree situational awareness, and MUCH faster reflexes than a human. So if there is a safe path to avoid a collision, and there usually is, FSD can find it.
I agree and have been asking for Lewis Hamilton and Max Verstappen to be the drivers to emulate
Yes, E2E networks can automagically figure out what to do with all the various sensor modalities, which conveniently avoids the need for hand-coded heuristics or even human intuitive understanding of the sensor data. But there are three major downsides to making a totally new sensor suite. The first is the cost and complexity to redesign the vehicle and rework the production lines. The second is that the existing fleet is now useless for gathering training data for the new fleet, since the sensors are completely different. And third is that making realistic synthetic data (to make up for downside #2) for multimodal sensor suites may be far more difficult than it is for pure vision.
Assume you’re talking about the future here.
Real-world human reaction time in emergency-braking scenarios, including the time required to move the foot from the accelerator to the brake and begin pressing it to slow the car, is around 1.5 seconds. Tesla's current 300-400ms is very much faster than that, and will continue to improve. I fully expect that Tesla will eventually implement a parallel "fast path" network for emergency collision avoidance, roughly akin to the human fast reaction (mediated by spinal cord neurons, not the brain) when e.g. touching a hot stove. Maybe HW5 will drop the latency in half. Maybe HW4 already does, once they start using it natively instead of using it to emulate the HW3 network, which is my understanding of how the current deployment works.
You’d need to demonstrate that. Humans are insanely fast.
FSD is 300-400ms from "photons in" to "control out". There's no further lag; if the "control out" tells the car to brake, the deceleration starts right away. FSD doesn't have a slow analog nerve impulse that has to travel from brain to foot, and it doesn't have a physical foot to move from accelerator to brake and start pressing it. Whatever time it takes a human to do all that, FSD doesn't have that lag. In emergency-braking scenarios, FSD’s reaction time (photons-to-deceleration) has been empirically measured in the real world at 300ms, whereas human+car photons-to-deceleration latency is more like 1500ms. That’s the apples-to-apples comparison.
I suspect this perception is mostly a training issue, not a response-time issue. I.e. the neural net seems slow because it doesn't think it's time to brake yet, perhaps because it's trained to simulate human response time for non-emergencies, not because it hasn't taken the input into account. With improved training, it may start braking sooner in the situations you think it should.
How are you measuring this? Again, I think the perception is because the neural net thinks it isn't time to act yet, not because the neural net hasn't fully processed the input. In non-emergency situations, it's probably simulating slow human reaction time, because that's what it's been trained to mimic.
Oh, no question v12.3.6 makes mistakes left and right. The architecture is undoubtedly capable of far higher reliability and correctness than it's currently achieving, and its current version makes a lot of bad decisions, including to not act when it should. My point is that when it does make a correct decision to act in an emergency situation, it can act (or to be more precise, it can cause the car to act) much faster than a human could.
There's no evidence here. I strongly doubt this.
Just based on disengaging routinely before the car does anything.
Never seen any evidence of this.
It’s way slower from noticing the yellow to applying the brakes than a human is even with identical reaction times. It just does not react quickly and eases it in over time even though it has perceived the light change. Not surprising of course.
