-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
It is not interesting to know where an object is, but more important is to know what an object is going to do so you anticipate accordingly
Agreed. This only helps in letting you know that you have detected all the objects in the scene.
EVNow
Well-Known Member
But easier than full robotaxi. I think cities will be ok with low speed (<35 mph) driving in urban/sub-urban roads. Tesla will need a small "rescue" team to rescue cars that are stuck during summon/park - but still better than a full time backup driver.
adiggs
Well-Known Member
This is a point where you and I see things differently. I see no functional difference in the ease of licensing between getting regulatory approval for an empty vehicle driving itself to a pickup point (at which point the human will sit behind the wheel and be responsible for accidents should once arise), and getting regulatory approval for FSD. Either way, you need regulatory approval for vehicles that move around on their own with no human in a backup position.
I don't see FSD while empty, but human having some backup responsibility while the car drives when there is a human in the car, as being a useful step closer to regulatory approval for FSD.
EVNow
Well-Known Member
The difference is not just in human being inside and no one inside, though I'm sure it would be considered. The bigger difference is in speed and thus routes allowed to take. For eg., IIRC Waymo and others got approved for low speed testing in CA first.
ps :
Currently CA has these 3 permits. I wonder whether Summon/Park would essentially be the second permit here, not the third.
- a testing permit, which requires a driver
- a driverless testing permit
- a deployment (public use) permit
Last edited:
Summon/ park is for use on private property only where traffic laws may not* apply (and it requires user input).
* depends on jurisdiction
EVNow
Well-Known Member
I can see Uber drivers doing this. They can make more money than being a Uber driver. May be this is the first step before driverless long distance Summon/Park feature is enabled.
EVNow
Well-Known Member
Yes - the one I'm suggesting is essentially summon/park over long distance.
Doggydogworld
Active Member
I don't see a 35 mph limit being useful. Uber killed Elaine Herzberg at 39 mph. If you're safe enough to drive around pedestrians you're safe enough to drive, period. As a practical matter, there are plenty of places you can't reach on 35 mph roads and lots of places you can't reach without crossing 35+ mph roads. You can't run a service in which 35 mph "islands" randomly lose service as cars move around.
I don't recall Google/Waymo ever being limited to low speeds. Their Firefly car was a NEV and thus limited to 25 mph (I think) in CA. But they also had full DOT-legal test vehicles at the time.
Last edited:
true, but it assists in correctly detecting where the objects are in the scene.
Vision is obviously the most important part of autonomous vehicles to get right. I walk my kids to school, there is a key intersection, we (my kids and I) interpret facial expressions (and gestures) of approaching drivers, and they interpret our expressions and intent. LIDAR and RADAR will not interpret human facial expression. LIDAR and RADAR may correctly locate where we are, and our vectors, but it will not provide input for interpreting facial intent, vision may. LIDAR and RADAR do not provide input from roadside instruction.
but a NN with vision only will self reinforce both good and bad interpretation.
There are multiple methods to estimate depth from a scene. But these are estimations, they may be correct 50%, 90% 99% 99.9% and so on. LIDAR and RADAR measure depth, which is different to estimate. A static vision scene will tend to use stadiametric rangefinding, that is, the-recollected height of an object has a direct, distance based correlation to the pixel count height of an object. Humans, dogs, deer, etc are all connected to the ground and can have their distance estimated. In my country there are kangaroos, kangaroos are not connected to the ground, which means that a vision system trained for objects connected to the ground will interpret a kangaroo very weirdly (literally as a giant in the distance, which zooms closer and shrinks, and then retreats and becomes a giant again). RADAR LIDAR can override the incorrect estimations, by direct measurement.
A moving vision scene can also use the relative change in size for further input for locations. Not strictly the same as stereoscopic, but in many ways comparable. If a stationary object grows of reduces by a certain percentage between images, its distance can be estimated.
Tesla is not vision only, they will use RADAR.
so what this comes down to is not whether vision only is the optimal way, but whether vision with RADAR is better than vision with LIDAR or is vision with both LIDAR and RADAR better, and much of this is determined by the actual RADAR/LIDAR sensor hardware/software.
which comes to the next issue, demonstration of reliability. If I want a single system which is reliable over 1,000,000 units (hours or km etc), there is a certain amount of time needed to validate that system. If I have 2 dependent systems, I now need twice the validation effort. But if I have 2 independent systems, I now only need the square root of the effort to validate.
I will also add a truism
The average driver is far safer than the average driver.
Which is to say, average driver statistics include those who have removed themselves from the gene-pool. Those of us who remain have our average driver results vastly dragged down. For instance, If I were to have a fatality, my future road insurance contributions are basically eliminated, and those costs must be borne by the other drivers who constitute the bulk of the driving population, and don't have fatalities themselves.
But that principle is applied even further, my location greatly determines what rate I would pay, because there needs to be a balance in money in/money out. Therefore if my location is associated with high costs for the insurer, then the others in the location are charged more, the local average subsidizes the local average.
The average driver is far safer than the average driver.
Which is to say, average driver statistics include those who have removed themselves from the gene-pool. Those of us who remain have our average driver results vastly dragged down. For instance, If I were to have a fatality, my future road insurance contributions are basically eliminated, and those costs must be borne by the other drivers who constitute the bulk of the driving population, and don't have fatalities themselves.
But that principle is applied even further, my location greatly determines what rate I would pay, because there needs to be a balance in money in/money out. Therefore if my location is associated with high costs for the insurer, then the others in the location are charged more, the local average subsidizes the local average.
The average driver is far safer than the average driver.
Fatality Facts
the average driver is not drunk, but the average fatality almost is drunk.
america has about 35,000 fatal road accidents per year, approximately 1:1 ratio fatal accident : fatality
about half of those fatalities are drivers(17,000), 11,000 drivers have a BAC reading and 5,000 have a BAC reading of >0.08
ok, so 35,000-11,000 is 24,000 fatal accidents involving sober drivers. lets assume none involve invasive movements from drunks, pedestrians and cyclists. so if america has 225,000,0000 license holders having 24,000 fatalities per year, thats almost a 1:10,000 ratio of fatality per year. so each year, 9,999 average drivers have no fatalities, 1 average driver has a fatality.
1:10,000 ratio of fatality per year , is high, over 50years thats 1 in 200 odds.
so 99.5% of drivers are safe enough to never ever have a fatality, its the last 0.5% that are the problem
having a car that is far worse than the average driver, but far better than the average drunk driver, could save many lives!!!
Fatality Facts
the average driver is not drunk, but the average fatality almost is drunk.
america has about 35,000 fatal road accidents per year, approximately 1:1 ratio fatal accident : fatality
about half of those fatalities are drivers(17,000), 11,000 drivers have a BAC reading and 5,000 have a BAC reading of >0.08
ok, so 35,000-11,000 is 24,000 fatal accidents involving sober drivers. lets assume none involve invasive movements from drunks, pedestrians and cyclists. so if america has 225,000,0000 license holders having 24,000 fatalities per year, thats almost a 1:10,000 ratio of fatality per year. so each year, 9,999 average drivers have no fatalities, 1 average driver has a fatality.
1:10,000 ratio of fatality per year , is high, over 50years thats 1 in 200 odds.
so 99.5% of drivers are safe enough to never ever have a fatality, its the last 0.5% that are the problem
having a car that is far worse than the average driver, but far better than the average drunk driver, could save many lives!!!
Doggydogworld
Active Member
It becomes a lot easier, both legally and practically, to force unsafe drivers to give up their keys if there's a comprehensive and affordable robotaxi service available. Presented properly, this could be a key selling point when seeking approval at the state and local level.
A reason for state approval, but a reason not to want to robotaxi your car. Puke takes a long time to get out.
Similar threads
- Question
