I recently went on a 4 hour drive that was mostly country back roads. Two hours of the trip took me through roads out in the middle of nowhere and for about an hour of the drive there were huge, gigantic potholes that looked big enough to severely damage a tire or shit maybe even cause more severe damage if you hit them just right.
I’m kind of wondering would an autonomously driving vehicle be capable of negotiating such terrain with ease or would that be a serious problem for even advanced software?
That is a situation that requires evaluation through experience and detail perception, and is the kind of edge case where the judgment a human driver will be needed to at least guide such a system for the foreseeable future. But the vast majority of driving occurs on paved roads that are generally maintained to some minimal standard, and while an autonomous piloting system would have to be able to recognize hazards and obstructions such as giant potholes, accidents, construction, pedestrians, et cetera it could navigate by avoiding the potential hazard or taking an alternate route.
I haven’t read all through this thread, but has anyone mentioned the fact that a person stepping out into the road in front of an electric computerised vehicle will bring it to a sudden stop, - great for pranksters & robbers!
That is both extremely uncreative of you as a criminal and speaks well of your law-abiding leanings.
A more motivated criminal might grease the road on the remote path of an unmanned delivery truck and ‘push’ it (glitch the sensors, hack the systems etc.) off the road and into a tree or a ditch. Then make off with the payload. Easy pickings and relatively low consequences if caught (theft not murder). There is a whole new world of possibilities.
Does this threat really need yet another list of every time Elon Musk has claimed or promised innovations in a timeline that turned out to be hopelessly optimistic?
Anecdotally, I’ll note that I’ve been run out of lane twice earlier this year by Teslas that were clearly on Autopilot failing to maintain landholding. In one case, the vehicle was ping-ponging between two lanes while the ‘driver’, oblivious to the world, was watching a video on his tablet. And this is not a singular observation; there are numerous online reports of the Tesla autopilot not being able to maintain stable lane position or spontaneously shutting down.
It the system can’t even hold to a single lane—something that conventional, non-AI driver assist systems can do reliably—then I’ll take Musk’s prediction of when autonomous piloting will be ready for wide deployment under advisement.
I ran into an entertaining situation driving in to work one time that made me wonder how an autonomous vehicle would handle it (probably would require some method to provide it external instructions):
I turned left at a light onto a small road next to my building (in between multiple buildings). There was a truck in my lane a little ways ahead stopped (nobody in it) and blocking traffic but my turn was before it.
I drove about 3 car lengths (with 3 cars following behind me doing the exact same thing) and started my right turn into the alley to get into the parking garage.
About 1/2 through the turn I stopped because there was a delivery truck backing up because he was blocked in the alley by a parked delivery truck. To my left (left side of the road) there was a delivery truck doing the same thing, for the same reason (the left alley was also blocked by an empty delivery truck), but we were blocking his ability to get completely out, so he was waiting.
At this point everyone recognized the situation and the simplest action was for my lane of cars to reverse and let the trucks in the alleys out and proceed, or wait for a while for the trucks blocking the alleys to move along.
Seems like the autonomous cars will need to have helper mode where they receive instructions on how to get around weird situations.
When AVs can manage the basics (like recognizing and being able to stay in their lane!) of driving on limited-access highways, I’ll start wondering how they can handle oddball situations like with the delivery truck in the alley. Until then, it’s like wondering how your six year old who has trouble adding and subtracting is going to manage Lebesgue integration.
I EMT’d for a rural ambulance service, paid by the run. We joked about greasing the Eight Ball Tavern’s steps to generate a bunch of nice fractures to transport.
Greasing roads to pick off AVs runs the risk of sluicing humans by mistake. A more nerdish approach employs a signal blanker. Scramble the wee little GPS brain, direct it down a nearby alley, block it in, and… no, I can’t advocate committing a crime. Forget I mentioned this. Put down that electromagnet. And the laser. And the Android phone. Slowly…
Tesla says they can approximate lidar with the visual cameras. If so, that’s freaking amazing.
A lot of improvements have been happening in lidar in rain and snow to make adjustments on the fly, but it’s ultimately still limited. I wonder if Tesla’s visual cameras can use the same adjustments?
It’s a fundamental principle of advertising that you can say anything you want as long as no one can directly disprove it.
I’ll wait to see an actual demonstration that Tesla can achieve the same fidelity in sensing the real world through visual information as it can via LIDAR, because that is something even human drivers cannot do and I have yet to see that computer vision systems have suddenly demonstrated superior capability to human vision.
Those two words are not the right way to start a serious discussion.
We need facts.
Tesla is famous for making proud claims of as-yet-unproven success. What I find interesting about this link is that the author mentions the location where the claim was announced: “the Scaled ML Conference”.
Now I have no idea what that is…but it sounds like a professional environment, where maybe some peer review is going on. If so…bring it on! I’d love to hear from the peers…with unbiased analysis.
They aren’t saying they can achieve the same fidelity, they are saying that their goal is to achieve a good enough approximation (of pixel depth) that they can use cheap cameras instead of expensive lidar.
It sounds like their presentation was more about progress towards that goal as opposed to claiming they’ve achieved it.
The new Tesla Model Z has 500-mile+ all-electric range, super-fast recharge time of under 5 minutes, Level 5 full automation, and a revolutionary new cloaking device that allows it to be completely invisible!!! Step right up, folks, to put down a deposit on this revolutionary new car! It’s fit for a king–or even an emperor!!!
I actually quite like Tesla–although I am very skeptical of near-term self-driving cars–and I kind of like Elon Musk more generally–although there’s no question he tends to over-promise, and often he can be a total asshole–but I just couldn’t resist.
Which could somehow do 0-60 in like 3 seconds. I also saw a Ferrari try to do that recently (not at the same time)-I am certain the Tesla would beat it.
Oh, to be clear, I don’t believe non-lidar approaches are even workable for standalone AV, at least not in our lifetimes. But it would be impressive to get that kind of fidelity, even in the clear daylight.
As for Tesla, there’ll also never be L5 in our lifetimes, regardless of what he promises.
If I didn’t already have AV cars as part of my newsfeed, I’d think Google news was reading my posts… In any case, this Forbes article regarding some new Tesla beta programming just popped up in my feed:
i wouldn’t say decades away … more likely, a handful of years at worst. when viewing the first video below … might be such a thing as everything was programmed into the navigation … not intended to be a true ‘depiction’. whereas, the second video seems totally ad-lib.
