-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FIRMWARE UPDATE! AP2 Local road driving...and holy crap
- Thread starter Skalinator
- Start date
Couldn't agree with you more. That's pretty much looks like he and Tesla operates. They mostly, if not always, miss deadlines, but in the end they are still beating their competitors (so far)...
kosarfan
Member
josher32
Member
Yes rumor our there is this update 17.7.2 adds parallel parking. I'm hoping mostly though it adds speed limit for AP on freeway!
17.7.2 downloaded - parallel parking
josher32
Member
henderrj
Active Member
Holy crap! I just spent 2 hours flipping through this thread.
sillydriver
Member
My concern is that we may be seeing evidence that good autopilot software has much more scarcity value than Tesla thinks. After all, why wouldn't MobilEye fear that cutting off a major customer like Tesla would worry and alienate their other customers? The answer may be that they know the hoops that they themselves had to jump through to get it to work reliably, so they believe Tesla will badly stumble when they try to recreate autopilot themselves. That would demonstrate to their other customers that MobilEye has something of great value, so that MobilEye holds the cards in their own negotiations. Tesla's stumble would demonstrate (for MobilEye) that it takes more than a car company's hiring some deep-learning PhDs to create autopilot software that works in enough real-world corner cases to be safely used on the road. Then the value of MobilEye as a company would rise.
I frankly hope what I've just said is not true!
I frankly hope what I've just said is not true!
Joe F
Disruption is hard.
I'm just surprised neither company uses map data to augment AP. While testing at the limited 50 MPH on a highway when the right solid lane line ended, the car did the seek for the cherished center of the lane, right into the path of an entering car in the acceleration lane of an on-ramp. Or at least it started to before I took control.
Map data would have corrected all of the diving for entrance and exit lane errors.
Map data would have corrected all of the diving for entrance and exit lane errors.
Yes, I didn't have that before this update either. It works fairly well as long as there are clear lane markers/curb on both sides (so not at intersections). It is still limited to 35mph.
lunitiks
Cool James & Black Teacher
Joe F
Disruption is hard.
Right. Where the density of Tesla's, over time, helps. Ideal design choice? The question is, what do the folks who have less density have to look forward to? Many years of traveling virgin roadways before the "fleet" of 1 gets the idea it shouldn't dive for an exit or entrance lane? 400K model 3's perhaps?
The map data has been out there for years. How does one defend the "Not Invented Here" approach?
There definitely is. As I drive in Finland I everyday use roads where line marks are covered with snow. Of course AP doesn't activate if there are no visible lines or lead car, but I every day encounter a situation, where I can engage AP with lead car or lines, but later lead car and/or lines disappear. AP1 still stays active and follows high precision map which it downloads from mothership. Of course problem is, that GPS's precision is not so good and I often need to take over, because car drifts off the line. But with good luck I can drive long without lead car and visible line markings.
I really don't think fleet learning is an acceptable fundamental approach to autonomous driving. Sure it's a viable fallback in the absence of other data, but we can't patch up every "lane lines are poor" issue with dead-reckoning based off crowdsourced GPS data and reliance on the accuracy of the GPS data at any given moment.
There needs to be more focus on driving heuristics that are NOT based off centering between two white lines: For example, basic computer vision detection of ditches / guard rails / grassy medians to understand where is clearly off limits… understanding of cars around you and keeping your space from them…. following wheel grooves already made...
When humans can't see lane lines, we follow a much fuzzier set of rules not necessarily based off where we remember/thought lane lines used to be. Right now, neither AP1 nor the early AP2 seems to have this kind of behavior programmed in.
There needs to be more focus on driving heuristics that are NOT based off centering between two white lines: For example, basic computer vision detection of ditches / guard rails / grassy medians to understand where is clearly off limits… understanding of cars around you and keeping your space from them…. following wheel grooves already made...
When humans can't see lane lines, we follow a much fuzzier set of rules not necessarily based off where we remember/thought lane lines used to be. Right now, neither AP1 nor the early AP2 seems to have this kind of behavior programmed in.
Google/Waymo produces a high resolution data intensive 3D map of the environment where their cars will drive. This needs to be done before the FSD car can be used. The Waymo FSD car doesn't make maps, but uses the motherships maps which have undoubtedly been massage to allow the car to determine its exact location.
Mobileye's claimed innovation is a low data rate technique for the cars themselves to make this map by uploading to the cloud. This is Musk's first pillar of fleet learning. AP1 drivers may have a some feel for if it works well enough to enable full automation.
For mapping Tesla has probably just swapped out Mobileye for nvidia. Both companies highlight their technology's ability to place the vehicle on the proper part of the roadway. But mobileye is a more mature and comprehensive solution than nvidia. Tesla has a lot of holes to fill to replace nvidia with tesla.
Mobileye's claimed innovation is a low data rate technique for the cars themselves to make this map by uploading to the cloud. This is Musk's first pillar of fleet learning. AP1 drivers may have a some feel for if it works well enough to enable full automation.
For mapping Tesla has probably just swapped out Mobileye for nvidia. Both companies highlight their technology's ability to place the vehicle on the proper part of the roadway. But mobileye is a more mature and comprehensive solution than nvidia. Tesla has a lot of holes to fill to replace nvidia with tesla.
Nigel Tufnel
Member
It could have been worse, a lot worse. You could've spent two hours listening to mis/disinformation from Tesla representatives on this subject. For example, how often do you think they would use the word "heuristics"?
Almighty Elon, grant me the strength to deal with your employees.
Similar threads
