There’s a HUMAN inside Tesla FSD 12.3

AI DRIVR
16 Mar 202428:35

Summary

TLDRThe video script offers an insightful analysis of Tesla's Full Self-Driving (FSD) technology, highlighting the significant improvements in version 12.3. The AI demonstrates a more human-like driving behavior, particularly in handling unprotected turns and maintaining consistent speed. Despite some critical disengagements due to road closure misunderstandings, the overall performance is promising, suggesting that with continued updates, FSD could become even more reliable and user-friendly.

Takeaways

  • 🚗 Version 12 of Full Self-Driving (FSD) technology shows significant improvement in driving like a human, with more natural and predictable behavior.
  • 🛣️ The car's performance in unmarked intersections and parking lots has greatly improved, showing cautious and patient behavior, similar to that of a human driver.
  • 🚶 The FSD beta effectively yields to pedestrians, demonstrating a good understanding of their movements and prioritizing safety.
  • 🚦 The vehicle handles unprotected turns with more caution, creeping out to ensure visibility of oncoming traffic before proceeding.
  • 🚗🔄 The car still faces challenges with road closures, resulting in critical disengagements when it attempts to navigate through them.
  • 🚘👨‍💼 There were instances where the car's automatic emergency braking system intervened, showing that safety systems are actively working to prevent potential accidents.
  • 📈 The overall speed maintenance and consistency have improved, making the ride smoother and more enjoyable.
  • 🚫 The car still requires human intervention for certain maneuvers, such as making a U-turn in a small space or navigating through road closures.
  • 📌 The FSD beta's ability to anticipate other drivers' actions and react smoothly to unexpected situations is a notable improvement.
  • 🚨 Despite the advancements, there are still areas that need refinement, particularly in understanding and reacting appropriately to road closure situations.
  • 🚀 With the rapid pace of improvement from version 12.2 to 12.3, expectations are high for the future capabilities and release of FSD technology.

Q & A

  • What is the main topic of the video transcript?

    -The main topic of the video transcript is the evaluation of Tesla's Full Self-Driving (FSD) version 12.3, focusing on its performance in various driving scenarios and comparing it to previous versions.

  • What are the two possible reasons the narrator considers for the slow rollout of version 12 of FSD?

    -The two possible reasons the narrator considers are that either the new approach using human data to train a generalized driving AI is working, or Tesla employees are remotely driving all the cars from their offices.

  • How does the narrator describe the parking lot performance of FSD version 12.3?

    -The narrator describes the parking lot performance as improved, with the FSD version 12.3 being able to handle complex situations like unmarked intersections and interacting with pedestrians more effectively than before.

  • What is the narrator's opinion on the potential release of Smart Summon in future FSD updates?

    -The narrator is excited about the possibility of Smart Summon being introduced in future FSD updates, as it could further enhance the convenience and capabilities of the self-driving system.

  • How does FSD version 12.3 handle unprotected turns compared to version 11?

    -FSD version 12.3 handles unprotected turns more cautiously and naturally, creeping forward to get a good view of incoming traffic before proceeding, whereas version 11 had more trouble with these types of turns.

  • What is the narrator's view on the AI's ability to handle a situation with a pedestrian crossing unexpectedly?

    -The narrator believes that the AI handled the unexpected pedestrian crossing situation better than a human driver might have, showing an improved understanding and reaction time compared to previous versions.

  • What are the critical disengagements mentioned in the transcript, and what was the common issue?

    -The critical disengagements were instances where the FSD system would have gotten into trouble without human intervention. The common issue was the system's inability to properly handle or understand road closures.

  • How does the narrator describe the automatic emergency braking system in FSD version 12.3?

    -The narrator mentions that the automatic emergency braking system in FSD version 12.3 seems to have been turned up, leading to an incident where the car skidded to a stop while trying to navigate a tight gap.

  • What is the narrator's overall impression of FSD version 12.3 after the test drive?

    -The narrator is impressed with the improvements in FSD version 12.3, particularly in maintaining consistent speed, handling unprotected turns more naturally, and exhibiting more human-like behavior. However, they also note the critical disengagements related to road closures.

  • What feature from previous FSD versions does the narrator miss in version 12.3?

    -The narrator misses the models for objects like trash cans and cones from previous FSD versions. In version 12.3, these have been replaced with blobs, which, while not as visually appealing, provide more information for the driver.

  • What does the narrator suggest as a potential future improvement for FSD?

    -The narrator suggests that the ability for FSD to reverse and perform actions like three-point turns on its own would be a significant future improvement, as it would resolve some of the non-critical disengagements experienced during the test drive.

Outlines

00:00

🚗 Introducing Version 12: A Leap Forward in Full Self-Driving

The video begins with a discussion on the new approach to training a generalized driving AI using human data, highlighting the improvements in Tesla's Full Self-Driving (FSD) version 12. The host shares their belief that this version feels the most human-like yet, with notable advancements in parking lot navigation and handling unmarked intersections. Excitement is expressed over the potential release of Smart Summon in future iOS app updates and the possible return of Auto Park. The video showcases how the car confidently handles complex situations such as navigating through a busy parking lot, dealing with pedestrians, and managing tight gaps, emphasizing the significant leap in capability compared to previous versions.

05:02

🌟 Version 12's Enhanced Perception and Decision Making

This paragraph delves into the enhanced perception capabilities of FSD version 12, noting the replacement of detailed models with 'blobs' for objects like trash cans and cones, which surprisingly offer more information. The host discusses the car's ability to handle complex situations such as road closures and unprotected turns, highlighting the car's cautious approach and improved decision-making. A particular focus is on the car's interaction with pedestrians and other vehicles, showcasing how the FSD beta can anticipate and react to their movements more effectively than earlier versions. The host also expresses gratitude towards Patreon supporters, attributing the improved video production quality to their contributions.

10:04

🛣️ Navigating Challenges: Road Closures and Construction Zones

The host discusses the challenges faced by FSD version 12 when dealing with road closures and construction zones. Despite the car's improved capabilities, it still struggles with understanding and reacting appropriately to road closures, leading to a few critical disengagements. The video provides examples of these situations, including the car's inability to perform a U-turn in a small space and its difficulty in rerouting around obstacles. However, the host remains optimistic about the future improvements, particularly once the car gains the ability to reverse on its own, which is expected to significantly enhance its navigational skills.

15:05

🚶‍♂️ Improved Pedestrian Interactions and Human-like Driving Behaviors

This section of the script focuses on the remarkable improvements in the car's interaction with pedestrians and its adoption of human-like driving behaviors. The host narrates instances where the car exhibits politeness and预判, such as allowing a car to merge or waiting for pedestrians to cross, even when they might not be expected. The video also highlights the car's ability to handle complex traffic situations, like navigating through a roundabout and making unprotected turns, with a level of caution and decision-making that closely mimics human drivers. The host expresses their amazement at the car's ability to read and respond to traffic signs, such as 'no right on red', and its overall progress in safety and predictability.

20:06

📈 Version 12.3: A Comprehensive Review and Future Expectations

In the concluding paragraph, the host provides a comprehensive review of FSD version 12.3, summarizing the significant improvements made over the previous version. They note the car's enhanced ability to maintain consistent speeds and handle unprotected turns more naturally, as well as its increased human-like behavior in traffic situations. The host acknowledges the three critical disengagements related to road closures and the four non-critical disengagements, suggesting that these issues are fixable and will likely be resolved with future updates. They express optimism about the rapid pace of improvement and the potential for full self-driving capabilities in the near future, while also contemplating the need for new ideas for their YouTube channel as the technology advances.

Mindmap

Keywords

💡Full Self-Driving (FSD)

Full Self-Driving (FSD) refers to the advanced driver assistance system developed by Tesla that aims to enable a vehicle to navigate and operate autonomously without any human intervention. In the context of the video, FSD is being evaluated based on its performance in various driving scenarios, showcasing its improvements in decision-making and handling complex traffic situations.

💡Human Data

Human data in the context of autonomous driving refers to the information collected from human drivers' behavior, which is used to train and improve the AI algorithms of self-driving systems. The video suggests that the use of human data has contributed to the more natural and human-like driving style exhibited by the FSD system.

💡Parking Lot Performance

Parking Lot Performance refers to the ability of a self-driving car to navigate, park, and exit parking lots, which is a complex task involving decision-making and interaction with other vehicles and pedestrians. In the video, the FSD's parking lot performance is highlighted as having significantly improved, showcasing its capability to handle unmarked intersections and tight spaces.

💡Unprotected Turns

Unprotected turns are maneuvers at intersections where there are no dedicated traffic signals or signs to control the flow of traffic, requiring the driver to make decisions based on the behavior of other vehicles and pedestrians. The video emphasizes the improved performance of FSD in executing unprotected turns, demonstrating its enhanced ability to assess traffic conditions and make safe decisions.

💡Critical Disengagements

Critical disengagements refer to instances where the self-driving system is unable to handle a situation and requires immediate human intervention to prevent potential accidents or issues. In the context of the video, critical disengagements highlight areas where FSD still needs improvement, particularly in understanding and reacting to road closures.

💡Non-Critical Disengagements

Non-critical disengagements are situations where the self-driving car disengages from autonomous mode, but the issue could be resolved without immediate human intervention or does not pose a direct safety risk. These instances indicate areas for refinement in the FSD system, such as navigating tight gaps or performing complex maneuvers like U-turns.

💡Automatic Emergency Braking

Automatic Emergency Braking is a safety feature in modern vehicles that uses sensors and cameras to detect potential collisions and applies the brakes automatically to prevent or mitigate crashes. In the video, FSD's automatic emergency braking is highlighted as an active safety system that intervenes in certain situations, showcasing its role in maintaining safe driving.

💡Visualization

In the context of self-driving technology, visualization refers to the graphical representation of the car's surroundings, including other vehicles, pedestrians, and obstacles, as seen through the sensors and cameras of the autonomous system. The video discusses how FSD's visualization has improved, now representing certain objects as 'blobs' for better understanding and decision-making.

💡Road Closures

Road closures are situations where a section of the road is blocked off, preventing vehicles from passing through. In the video, road closures present a challenge for FSD, as it struggles to understand and react appropriately to these scenarios, leading to critical disengagements.

💡Improvement Rate

Improvement rate refers to the speed and extent of advancements or progress in a technology or system over time. In the context of the video, the improvement rate of FSD is discussed in terms of the rapid enhancements made from one version to the next, indicating the potential for significant progress in the near future.

Highlights

The new approach using human data to train a generalized driving AI seems to be working, making the full self-driving (FSD) version 12 feel more human-like in its driving behavior.

FSD version 12 has shown improvements in parking lot performance and handling unmarked intersections, showcasing its ability to wait out complex situations with multiple actors involved.

Leaks in the iOS app updates suggest that Smart Summon may be coming soon, which would be a significant addition to the FSD features.

The FSD beta has demonstrated a more natural and smooth driving manner, avoiding the slow creep that signals intent to other drivers, thus feeling more competent.

Pedestrian interactions have seen considerable improvement, with the FSD beta yielding well for pedestrians, even in situations where the human driver might not have been as accommodating.

The FSD beta managed a complex maneuver of crossing multiple lanes to make a left turn at a light, showcasing its understanding of traffic patterns and lane discipline.

There's a notable difference in the way version 12 handles right turns at lights, with the system displaying more caution and better anticipation of traffic conditions.

The FSD beta's ability to handle unprotected turns has improved, with more cautious and human-like behavior, making it a safer and more predictable driver.

The beta's reaction to a pedestrian crossing in front of the car was well-handled, even when the human driver did not see the pedestrian until editing the video.

The FSD beta's automatic emergency braking system was triggered during a tight gap maneuver, demonstrating the safety systems' vigilance despite the beta's capability to handle such situations.

The beta's visualization system has replaced models for objects like trash cans and cones with blobs, which, while less aesthetically pleasing, provide more useful information for understanding the car's surroundings.

The FSD beta exhibited a better understanding of road closures and navigated more cautiously, although it still struggled with certain road closure scenarios.

The beta's ability to anticipate and smoothly handle the movements of other vehicles, even when they are indecisive or unexpected, showcases its advanced understanding of traffic dynamics.

The FSD beta's handling of a tough unprotected left turn, with its cautious creeping behavior and good visibility management, demonstrates a significant improvement over previous versions.

The beta's decision-making during a yellow light, choosing to accelerate through the intersection after initially braking, shows its capability to make quick, human-like judgments.

The FSD beta's navigation and driving behavior, particularly in complex situations like road closures and construction zones, indicate that while it has made great strides, there is still room for improvement.

The overall improvements from version 12.2 to 12.3 are substantial, suggesting that if this rate of enhancement continues, the future of full self-driving looks promising.

Transcripts

00:00

after watching this drive through San

00:01

Jose I think you'll agree one of two

00:04

things about full self-driving is true

00:06

either the new approach using human data

00:08

to train a generalized driving AI is

00:10

working or the reason version 12 has

00:12

been so slow to roll out is because

00:15

Tesla employees are the ones remote

00:17

driving all the cars from their offices

00:19

in all seriousness this really is the

00:21

most human feeling version of FSD yet

00:23

and there's some moments in this drive

00:24

you're going to have to see to believe

00:26

the first thing it's going to have to do

00:28

is escape this parking lot you can see

00:30

it pull up to this intersection pretty

00:31

cautiously and there's an approaching

00:33

car from our leftand side and also

00:35

someone leaving their parking spot from

00:37

the right hand side so it does decide to

00:39

wait it out parking lot performance in

00:41

general and especially the performance

00:44

in these unmarked intersections has

00:46

gotten really really good and I've also

00:49

seen some leaks in the IOS app updates

00:51

where it looks like actually smart

00:53

summon could be coming very soon which I

00:56

am super excited to try out and I'm also

00:59

excited that we're probably going to get

01:00

Auto Park back too you can see another

01:03

driver ahead of us leaving their parking

01:04

spot and what I love here is that it's

01:06

not coming to a full stop and doing a

01:08

slow 1 or 2 mph creep the entire time

01:12

and I know that may not seem like much

01:14

but creeping slowly really does show a

01:16

lot of intent to other drivers which is

01:18

what they're used to and makes it feel a

01:20

lot more competent there's a pedestrian

01:22

on the right hand side who just walks

01:24

out without a care in the world she

01:26

never looked back or anything I think

01:28

don't think she has a clue where here

01:30

but uh the beta does a great job

01:32

yielding for her I know there's a lot of

01:34

humans out there that wouldn't have been

01:36

quite as nice in this situation the exit

01:39

of this parking lot is unfortunately one

01:42

that is very common around here but one

01:44

that I do not like which is this

01:46

intersection right where the entrance is

01:49

it could be pretty tough to escape a

01:51

situation like this but luckily there's

01:52

a pedestrian who's walking across in

01:55

front of the car to the left and it

01:56

actually takes this opportunity and goes

01:59

for it this was extremely well done it

02:01

CH a perfect moment to pull out and this

02:03

is something that version 11 had quite a

02:06

bit of trouble with it would have been

02:07

yielding to pretty much every vehicle in

02:09

the scene now as we approach this

02:11

unprotected turn you may notice that our

02:14

leftand turn signal is on for this right

02:16

turn you can see on the navigation it

02:19

definitely knows that it has to make a

02:21

right turn but yet it has the left turn

02:23

signal on that is because we have to

02:25

make a left turn at the light which

02:27

means as soon as we get out of this

02:28

parking lot we're going to have to cross

02:30

over multiple Lanes to get all the way

02:32

over into that leftand turn lane and as

02:34

soon as traffic clears it goes ahead and

02:37

makes its move and Crosses all four

02:40

lanes at once and gets us all the way

02:43

into the leftand turn lane very well

02:46

done and I know there is going to be

02:47

some complaints about it crossing a

02:49

solid white line which we all know is

02:51

punishable by death in all 50 states so

02:54

go ahead and cry about it in the

02:55

comments now you may be able to just

02:57

barely make out this pedestrian on the

02:59

visual visualizations that I in the car

03:01

did not see at all so I thought this

03:03

person turning right was just stopping

03:05

in the road for absolutely no reason

03:08

here I did not see The Pedestrian

03:10

Crossing in front of him at all this is

03:13

one of those scenarios where I feel like

03:15

the car may have had a little bit better

03:17

of an understanding of what was going on

03:19

than I did in the car because I didn't

03:21

even know that pedestrian was there

03:22

until I started editing this video and

03:25

while that maneuver was very well

03:27

handled and very smooth I might add the

03:30

next one was definitely not we have a

03:34

right turn to make it the light ahead

03:35

and you can see a human driver squeezing

03:37

the gap between the trash can and the

03:39

car next to it and I could not believe

03:42

autopilot was about to try the same

03:43

thing but it did I'm going to let you

03:45

watch what happens

03:47

here oh my

03:50

God it actually triggers the car's

03:53

automatic emergency breaking and makes

03:55

it skid to a stop this is definitely

03:57

worth taking a closer look at that is

03:59

definitely a first for the channel you

04:01

can see that the Gap is pretty tight

04:03

here but we have seen version 12 shoot

04:05

tighter gaps in the past but I think it

04:07

was because it was heading towards the

04:09

SUV to the left as it was passing the

04:11

trash can that triggered it now you

04:13

might not know this but there was

04:14

actually another incident on version 12

04:16

that somebody recorded a it tapping

04:18

another car's bumper in a parking lot so

04:20

I have a feeling that may be the reason

04:23

that the automatic emergency braking has

04:25

been turned up to 11 here now to be

04:28

honest I'm not really sure what to file

04:30

this disengagement under I'm going to go

04:32

ahead and call it a non-critical

04:34

disengagement because to be honest I

04:36

think the car had it handled I think it

04:38

would have completed that maneuver just

04:41

fine but it was the automatic emergency

04:43

systems that overrode its decision and

04:46

brought it to an abrupt stop I seem to

04:48

believe in version 12's abilities even

04:51

more than the car does we do go ahead

04:53

and pull off to the side of the road

04:55

because we have reached our destination

04:57

and I go ahead and put another one in

04:59

and re-engage while the car is doing its

05:02

best to try to get around this trash can

05:04

here I think it's interesting to point

05:06

out that the models for a lot of things

05:08

including trash cans and cones and all

05:11

that are completely gone with version 12

05:14

you can see they've been replaced with

05:15

kind of just blobs that show up on the

05:17

visualization and although the blobs

05:19

might look a little bit ugly I think

05:21

they actually give you a little bit more

05:23

information and I'm hoping that they

05:25

take over the entire driving

05:27

visualization someday now we we have a

05:30

left turn to make here but as you may be

05:32

able to see off to the left the road is

05:34

closed there is enough room to get

05:36

around those cones though so let's see

05:39

what it wants to do here you can see a

05:43

very slow creep into the intersection I

05:45

almost feel like it knows it shouldn't

05:48

turn left here but it's still trying to

05:50

figure out a way to do it and it sits

05:52

here for quite a while until a car

05:55

approaches from behind us and at that

05:56

point it goes for it which does cause

05:59

our first critical disengagement because

06:02

unlike the last video where the road

06:05

closure said Road closed except for

06:07

local traffic this one is a real road

06:10

closure and I did not want to interrupt

06:13

whatever was going on over there so

06:15

looks like it's still having a bit of

06:17

trouble in those road closure situations

06:20

it's funny in all the road closures I

06:23

put it in it almost feels like it knows

06:25

it's not the right thing to do but

06:27

always ends up just trying it anyway

06:29

ways we're actually going to try to

06:31

navigate through that exact same road

06:32

closure later in the drive to see if the

06:34

same outcome happens and as we fast

06:37

forward to the next moment of Interest I

06:38

wanted to give a quick shout out to my

06:41

patreon supporters cuz I saved up every

06:43

dollar you guys have sent for months and

06:45

months and used it to buy a proper

06:48

editing computer which makes making

06:50

videos like this a lot easier cuz I

06:52

really was maxing out my previous

06:55

computer making these videos so thank

06:57

you for directly supporting the channel

06:58

and always know that every dollar I get

07:00

from patreon directly goes into

07:02

improving quality on the channel the

07:04

driver in front of us is a little bit

07:06

indecisive about his right turn up here

07:08

and you can see he slows down and then

07:10

speeds up and ends up making the ride at

07:12

a later time but full self-driving is

07:15

able to remain extremely smooth through

07:17

all of this it doesn't even slow down as

07:19

he's making the next right version 12

07:21

seems to be able to anticipate what

07:23

people are going to do really really

07:26

well and even when they do something

07:28

unexpected it's still still remain

07:30

smooth another really good example of

07:32

that up here we're on a pretty narrow

07:34

road where only one car fits at a time

07:36

and you can see it give the rideway to

07:38

this car but only slows down a little

07:40

bit and just maintains speed as it moves

07:42

over and then continues right on going

07:46

also love how early it offsets for

07:47

pedestrians compared to version 11

07:50

giving them no doubt that we see them

07:52

and we're controlling for them and there

07:54

was one specific really incredible

07:57

pedestrian interaction that happens a

07:58

little bit later in the video kind of a

08:01

funky intersection here where we have to

08:03

make this right turn and then make the

08:05

immediate left on that street that you

08:06

can see there and it gets a little bit

08:08

tripped up by this incoming car but as

08:11

soon as it was clear that he was going

08:12

to stop it goes ahead and proceeds no

08:15

problem we are about to reach our

08:18

destination and this time it actually

08:20

does not pull over it just kind of stops

08:23

in the road and it was fine cuz there

08:24

was nobody behind us so I just went

08:26

ahead and put in a new navigation route

08:27

real quick but kind of curious I don't

08:30

know the reason why sometimes it hunts

08:32

for a spot to pull over and other times

08:35

it just kind of stops in the road and

08:37

although it's doing a really good job on

08:39

these streets with no Lane lines or

08:40

anything going to go ahead and fast

08:43

forward until we get some more

08:44

interesting scenarios cuz there's a lot

08:46

we got to see in this video coming up

08:48

here we have our first pretty tough

08:50

unprotected left turn and the speed

08:52

limit on this road is 40 but as you can

08:54

see by the model y passing in front of

08:55

us traffic is Flowing a little bit

08:57

faster than that it does have to create

08:59

creep really far out here even over the

09:01

crosswalk in order to have any

09:03

visibility of incoming traffic and after

09:05

the car passes that comes from the right

09:08

it creeps out a little bit more make

09:10

sure it's safe and then proceeds you can

09:12

see it crossing right over those Lan

09:14

lines there taking the human path

09:16

through that turn for sure now it did

09:19

get pretty lucky in that unprotected

09:20

turn because there wasn't a lot of

09:22

traffic but I think unprotected turns

09:24

have actually been improved quite a bit

09:26

this version seems to be much more

09:28

cautious at intersection like this that

09:30

are kind of blind because of traffic

09:32

parked on both sides there was no creep

09:34

wall displayed in the visualizations

09:36

there but it did an excellent job of

09:38

creeping out right to the Limit before

09:40

proceeding when it had a good view on

09:42

both sides intersections like this are

09:44

pretty dangerous and unfortunately

09:47

extremely common around here but again

09:50

you can see how it creeps very very

09:52

slowly and kind of does one final stop

09:55

at the very end here before proceeding

09:58

confidently this seem to be pretty

10:00

consistent and it's nice because this is

10:02

exactly how humans behave in

10:04

intersections like this we creep out to

10:06

just before the point where we're

10:08

starting to obstruct traffic and then

10:09

take a look and then proceed when it's

10:11

safe there's a bus in our lane and

10:14

instead of staying behind it it actually

10:15

moves over to get beside it which is

10:18

probably the same thing I would have

10:19

done but there's actually a car ahead of

10:21

the bus and an immediate merge for a

10:24

moment it almost felt like it was

10:25

slowing to let the car in in front of us

10:27

but when they were slow to proceed off

10:29

the the light it just went for it

10:31

probably the same thing I would have

10:32

done in that situation and again loving

10:35

the way it's ignoring the lane lines

10:37

here and Crossing them very early you

10:40

can see it's uh pretty patient for this

10:42

unprotected left turn here I think the

10:44

last version would have just went for it

10:46

right now but this version stays a

10:48

little bit more patient and lets the

10:49

next car pass it does seem like they've

10:52

dialed back the assertiveness of these

10:54

unprotected turns which is probably a

10:56

good thing I must say that the amount of

10:58

progress I'm seeing in areas like those

11:00

compared to just one version ago is

11:03

really incredible and if these are the

11:05

kind of improvements we can expect going

11:07

forward 2024 is going to be a wild year

11:12

for full self-driving coming up here is

11:15

one more example of its creeping

11:16

Behavior at this uncontrolled

11:18

intersection with no signs or anything

11:20

you can see it just goes right to the

11:22

point where traffic is approaching slows

11:24

down and then proceeds once it can

11:26

actually see and we have a right to turn

11:30

coming up and based on our previous

11:32

experiences you'd think it'd be going

11:34

for it right now there's a huge gap and

11:36

it can easily fit in there but it does

11:39

not I think it probably learned its

11:42

lesson from the last time it tried to do

11:44

that I'm just kidding that's not a real

11:46

thing but that is one of the side

11:48

effects of having it completely AI based

11:51

is that you never really know what it's

11:54

going to do in any situation it really

11:56

makes things pretty interesting there is

11:58

a cyc to our right in an oncoming car

12:01

approaching and it decides to shoot the

12:03

Gap even running over the center line a

12:06

little bit which was extremely

12:08

human-like and probably how everybody

12:10

would have treated that situation that

12:12

is an area where older versions

12:13

struggled quite a bit now we've seen

12:16

quite a few pretty humanlike moves out

12:18

of the beta so far but this next one

12:21

definitely tops it it's probably one of

12:23

the most human things I've ever seen it

12:25

do we have a right turn to make at this

12:28

St light ahead and traffic in front of

12:31

us begins slowing down and they actually

12:33

come to a full stop here and a car

12:36

approaches from the right hand side out

12:37

of that parking lot instead of pulling

12:39

up behind the car in front of us which

12:41

would have blocked this guy into the

12:42

parking lot it takes the polite approach

12:45

and lets this guy go he even waves

12:47

thanks before he goes this was an

12:49

absolutely incredible move I remember

12:52

not that long ago full self-driving

12:54

wouldn't even drive over these green

12:57

markings on the road because it was

12:58

scary of them look how far we have come

13:02

and when I posted that clip on X there

13:04

were a few people that said the car

13:05

should have just taken the ride of way

13:07

which would have mean blocking him in

13:09

there and usually I am on team just take

13:11

your ride of way and don't be polite but

13:13

in that scenario I don't think the car

13:15

did anything wrong at all and I think

13:17

that was the correct move now our

13:20

navigation point is right up here off to

13:22

the left and about 100 ft and it does

13:25

seem like it just wants to come to a

13:28

stop in the middle of the road here

13:30

instead of trying to pull over again not

13:33

really sure when it decides to pull over

13:36

versus when it just stops in the middle

13:37

of the street but interestingly after it

13:39

sees this car approach it decides to

13:42

start moving again that is something

13:44

I've noticed with version 12 in general

13:47

is that if there's another car in the

13:48

scene it seems to make it kind of behave

13:51

differently like when we were at that

13:53

construction zone and a car pulled up

13:55

behind it it made it move and that is

13:57

something that that's pretty consistent

13:59

version 12 seems to have a little bit of

14:01

a sense of urgency I guess you could

14:03

call it when there's other traffic

14:05

around uh does end up just coming to a

14:07

complete stop here and unfortunately

14:11

it's brought itself down a dead end road

14:14

so after choosing a new navigation route

14:16

you can see it has to do a uturn in this

14:20

little court at the end of the road here

14:22

now version 12 can actually do u-turns

14:25

but it cannot back up so let's see if

14:27

it's able to complete the turn in this

14:29

small space starting out strong swinging

14:32

the wheel to the right but doesn't quite

14:35

turn hard enough to the left in order to

14:37

complete the maneuver and ends up just

14:40

stopping here which does cause our

14:42

second non-critical disengagement of the

14:45

drive cuz it needed a little bit of

14:48

human help at least until it has the

14:50

ability to reverse on its own and I did

14:53

only back up a little tiny bit here to

14:55

see if it could make this harsh turn by

14:57

itself and after re-engaging it actually

15:01

does completely turn the steering wheel

15:03

to almost the fully locked position to

15:05

complete this turn there's also a model

15:08

3 backing out of their driveway here and

15:11

just like my initial review of version

15:13

12 I just really love how much room it

15:16

gives people who are backing up and

15:18

doing Maneuvers like this makes them

15:20

feel very comfortable instead of getting

15:23

right up on them like version 11 used to

15:26

something else I should mention that's

15:27

been a pretty significant upgrade

15:29

compared to the first version of version

15:31

12 that I tried is that if you remember

15:33

I complained about the automatic speed

15:36

control setting because it would kind of

15:38

Go Fast and then slow down and speed up

15:41

again it wouldn't be able to maintain a

15:42

consistent speed and I'm happy to report

15:45

that for the most part this version

15:48

pretty much completely fixes it for the

15:50

most part it's able to stay very very

15:52

smooth we have a right turn to make at

15:54

the light ahead and again there's plenty

15:56

of room to go right by the SUV on the

15:58

right hand side but it does not go for

16:01

it I think it's still a little bit

16:03

scared because of what happened

16:05

previously but then it did the most

16:07

human thing I think I've ever seen it do

16:09

where the light turns green and the

16:12

person in front of us does not

16:13

immediately go they're kind of asleep at

16:15

the wheel and autopilot starts creeping

16:17

towards them like hey buddy are you

16:19

going to go or what absolutely

16:22

unbelievable undeniable human behavior

16:25

there that is definitely not something

16:27

that can be manually programmed into the

16:29

car in any meaningful way I can

16:32

understand the argument for people who

16:33

don't like that kind of behavior but I

16:35

really don't mind showing intention to

16:37

other drivers he may have saw us

16:38

creeping up behind him before he saw the

16:40

green light I have it navigating

16:42

straight through the construction zone

16:43

that tripped it up before but

16:45

interestingly instead of trying to go

16:47

around it this time it makes a right

16:49

turn instead now most of the time this

16:51

would have been a good thing because it

16:53

can reroute a different way but it

16:55

actually just turned into a dead end you

16:58

can see on the navigation now it says it

17:00

wants to make a uturn which there is no

17:04

possible way it's going to do in this

17:06

little section right here which does

17:09

cause our third non-critical

17:11

disengagement of the drive because this

17:13

needs a little bit of a human touch at

17:15

least until it has the capability of

17:17

backing up by itself I think things are

17:20

going to get super wild once uh full

17:23

self-driving has the capability of

17:25

reversing all on its own I'm secretly

17:27

hoping that that comes in the same

17:29

update as the actually smart summon

17:31

after re-engaging here we still have the

17:33

the same navigation route which means

17:35

we're making an immediate right here

17:36

into the construction zone but

17:39

unfortunately this time it does not

17:41

reroute on its own and it seems like it

17:43

wants to head straight into this area

17:46

causing our second critical

17:47

disengagement of the drive so looks like

17:50

for now at least version 12 is still

17:52

lacking the data or training it needs to

17:55

know what to do in a road closure

17:57

situation which I'm a bit thankful for

17:59

because that means this channel is still

18:01

useful coming up here there's a car

18:04

who's pulling out off the side of the

18:06

road towards the stop sign super super

18:09

slowly not sure what they're doing and

18:11

they actually move back over to the

18:12

right hand side of the road and you can

18:14

see we pass them with confidence now

18:17

although situations like that look

18:19

pretty mundane these are the types of

18:21

situations that normal self-driving cars

18:24

have a lot of trouble with because it

18:26

takes real intelligence to know that you

18:29

should pass in that situation cuz in any

18:31

coded sense that was a slow-moving car

18:33

that moved slowly towards the stop sign

18:35

and then stopped at it so we really

18:36

shouldn't pass but you know of course we

18:38

really should have one of the benefits

18:40

of being trained on human data you can

18:42

see a no right turn on red sign here and

18:45

for the first time ever it really does

18:48

seem like full self-driving is reading

18:50

the sign maybe reading is the wrong word

18:53

but it understands no creeping

18:56

whatsoever and just waits for the light

18:58

to turn turn green this is the only no

19:00

right on red light that I've been to so

19:02

far so I can't say for sure that

19:04

everywhere is completely fixed but it

19:06

used to have problems at lights like

19:08

those that didn't have a dedicated red

19:11

arrow to tell you that it wasn't okay to

19:12

make a right and that is progress

19:15

something interesting happens up here

19:17

where the navigation route says that

19:19

we're going to turn right in a few

19:21

hundred feet up here but instead the car

19:23

ends up just pulling over instead and at

19:26

first this kind of confused me I was

19:28

like why the heck is is it pulling off

19:29

to the side of the road here when it's

19:30

not done navigating but then I actually

19:33

looked at the navigation route that I

19:35

had Chosen and it pulled over at the pin

19:38

as you can see here so where it's pulled

19:40

over right where the pin was dropped and

19:42

completely ignored the actual navigation

19:45

instructions that's actually really

19:47

interesting Behavior if you think about

19:48

it because that means there's at least a

19:50

little bit of a disconnect between the

19:52

navigation system and full self-driving

19:54

it really did just pull over at the pin

19:57

and ignored the navigation route some

19:59

strategically placed trash cans up here

20:01

giving the beta a little bit more of a

20:03

challenge but handles it absolutely

20:06

beautifully with zero hesitation if you

20:09

look further down the road here you can

20:10

actually see there's a dog crossing the

20:12

street and interestingly it's visualized

20:14

on the autopilot display is a little

20:16

blob crossing the street and it also

20:18

starts offsetting for it a little bit

20:20

and moving to the left to give them more

20:21

room I know it's pretty hard to see but

20:24

if you go back I think it's actually

20:25

pretty incredible how early it started

20:27

rendering that dog on the display it saw

20:29

it from a ways out another unprotected

20:31

turn here where it has to creep quite a

20:34

bit out in order to have good visibility

20:36

of incoming traffic from the left and

20:38

this is an area where the last version

20:40

of version 12 was very aggressive but

20:43

again this one does a much nicer job and

20:46

creeps way slower until it can actually

20:48

see coming up here the car does

20:51

something really incredible as we

20:53

approach this next red light you might

20:55

be able to see a mother walking her

20:58

children across the street to the left

21:00

and instead of proceeding all the way to

21:02

the stop line full self-driving actually

21:05

stops a little bit early to give them

21:07

extra room this is just such nice

21:10

driving unfortunately the human to the

21:12

right doesn't do this and watch the

21:13

mother look at this car like hey man are

21:16

you going to stop or what and then after

21:18

they're completely out of our path it

21:20

goes ahead and moves towards the stop

21:22

line absolutely incredible I've seen

21:25

some people wonder how full self-driving

21:27

is ever going to be safer than a human

21:29

if it's only trained on human data and

21:30

it's simple it's just going to be very

21:33

good driving human data and not data

21:36

from that guy on our right as we

21:39

approach this next unprotected left turn

21:41

for some reason the car felt extremely

21:44

hesitant to proceed and I was like why

21:46

the heck is it so hesitant here and then

21:48

I saw The Pedestrian walking through the

21:51

crosswalk now you can go back and watch

21:53

that but this pedestrian seemed to

21:54

appear out of absolutely nowhere now you

21:57

can call me a bad driver all you want

22:00

but the car definitely handled that

22:02

particular situation better than I would

22:04

have my focus was on traffic ahead I

22:07

definitely would have saw him but I

22:09

would have made a much more awkward

22:10

situation for that pedestrian another

22:13

borderline two human of a move up here

22:16

where this light turns yellow and its

22:18

first reaction is to break then decides

22:21

nah and just guns it through the

22:23

intersection actually accelerating

22:25

through the yellow light here very very

22:28

humanlike thing to do and it's really

22:30

interesting watching the car think

22:32

through situations that need to be made

22:33

very quickly the car in front of us is

22:36

making a left-hand turn but there's

22:38

traffic stopped in front of him and at

22:39

first I was like wow he really obeys the

22:43

lane lines that are painted on the

22:44

ground but that's actually not what was

22:46

happening and he's actually turning left

22:49

into the driveway short of that so there

22:51

are a couple of impatient humans that

22:53

pass from behind but as soon as it's

22:55

clear we go ahead and pass as well so

22:57

far even though we've had a couple of

22:59

critical disengagements due to that

23:01

construction zone my God does this

23:05

version feel very very good and ready

23:08

for wide release I'm really expecting it

23:10

to be pushed out wider any day now

23:12

slowing this down cuz it's the only uh

23:15

roundabout that we went through it's a

23:17

very boring roundabout and it handles it

23:19

no problem at all but it is the only

23:21

roundabout of this drive so I had to

23:23

slow it down and evil me navigated into

23:26

that same construction zone as before

23:29

except from a different area this time

23:32

this time the road is completely blocked

23:34

off though with a gentleman standing

23:36

there telling us to stop and the car

23:37

stops very early here and doesn't

23:40

proceed forward and I think it's fair to

23:43

call this one another non-critical

23:44

disengagement since it stopped

23:46

completely by itself and didn't seem

23:48

like it was trying to proceed at all

23:50

doing a quick three-point turn here and

23:53

navigating to our last destination of

23:55

the day which is more in the downtown

23:57

area of of San Jose the road ahead is a

24:00

little bit confusing you can see we have

24:02

a right turn to make in a few hundred

24:04

feet but the road doesn't look like it

24:07

goes that way it sweeps off to the left

24:10

and then after we make the left the road

24:12

markings all just kind of disappear and

24:15

even though I think this is a pretty

24:17

confusing place to drive my goodness how

24:20

smooth the steering wheel stays through

24:22

the entire thing I would love somebody

24:25

to take version 11 through here and just

24:27

record it so we can have a comparison we

24:30

have another unprotected left turnup

24:32

here across a few lanes of traffic and

24:35

you can see after coming to a complete

24:36

stop it again slowly creeps forward

24:39

right to the point where it can see

24:40

incoming traffic waits for a good moment

24:43

to go and just goes couldn't have done

24:46

it better myself all the unprotected

24:49

turns I've had so far at least have been

24:52

considerably better even compared to

24:54

just one version ago I think we're going

24:57

to have to see Chuck unprotected left

24:59

turn which is super intense in order to

25:01

get a real idea of how good this thing

25:03

is but for me so far so good it has

25:07

nailed absolutely every single one of

25:09

them and seems to stop at a really good

25:11

spot in order to get good visibility of

25:13

incoming traffic from both sides we have

25:16

a left turn coming up here it's a little

25:18

bit past the truck and interestingly it

25:20

seems like it kind of knows this and

25:21

gets prepared to pass but it didn't end

25:23

up needing to as he made the turn now

25:27

unfortunately there is a sign in the

25:29

middle of the road that's saying that

25:31

the road is closed except for local

25:33

traffic H I know this got me in trouble

25:36

in the comments last time but since our

25:39

navigation pin was right at the end of

25:41

this street right off to the right I

25:43

just let it happen again technically

25:45

we're local traffic right but

25:47

unfortunately for full self-driving the

25:49

destination right off to the right up

25:51

here is beyond another road closure and

25:55

the street ahead of us is really closed

25:58

not just for local traffic only this is

26:01

a truly closed road you can see there is

26:04

a gap off to the left so the cones don't

26:06

completely block the road here and I was

26:08

hoping that it would just stop but

26:10

unfortunately it did try to pass off to

26:13

the left which did cause our third and

26:16

final critical disengagement and because

26:19

we've been driving for over an hour and

26:21

a half now and all my camera batteries

26:23

were dying at once this is where I

26:25

decided to call it in summary there's a

26:28

lot of things to be really excited about

26:30

with version 12.3 one of my biggest

26:33

complaints about it not being able to

26:35

maintain speed has been improved in a

26:37

big way and it finally feels like the

26:39

car can carry a more consistent speed

26:41

and it isn't slowing down unnecessarily

26:43

nearly as much as it used to and it

26:45

finally feels usable and comfortable on

26:47

wide open roads with no lead car in

26:49

front the behavior for unprotected turns

26:52

seems to be improved in a big way as

26:54

well with the car creeping in a much

26:55

more natural way and only proceeding

26:57

when it has has good visibility of

26:59

incoming traffic on both sides and For

27:01

Better or Worse it's more human than

27:04

ever doing things I'd never thought I'd

27:07

see the full self-driving beta do I'm

27:09

sure opinions on this will be split on

27:11

if this is a good or a bad thing but I

27:14

think you already know my opinion I

27:16

think if you're sharing the roads with

27:17

humans the best thing to do is be

27:19

humanlike and predictable and last but

27:22

not least all these improvements

27:23

happened on a single dot update from

27:26

version 12.2 to version 12.3 if this

27:30

rate of improvement continues I think

27:32

I'm going to have to think of some new

27:34

ideas for a YouTube channel and on to

27:37

the negatives there's no way around it

27:39

we had three what I would consider

27:41

critical disengagements where the car

27:42

would have gotten itself into some

27:44

trouble if I hadn't have intervened the

27:46

good news here is that they were all for

27:48

the exact same thing which is not

27:49

knowing what to do for a road closure

27:51

which seems like a fixable problem we

27:54

also had four non-critical

27:55

disengagements where the car needed a

27:57

bit of help on the first one the safety

28:00

systems were actually what disengaged

28:02

full self-driving and made me take over

28:04

immediately but honestly I think it

28:06

probably would have handled the maneuver

28:07

fine even if it was a smaller Gap than

28:10

most drivers would go for the other

28:12

three non-critical disengagements are

28:14

pretty easily solved once the car has

28:16

the ability to back up which I believe

28:18

we're going to be seeing very soon if

28:20

not on full self-driving at least summon

28:23

and I bet it won't be long after that

28:24

where the car will do it all on its own

28:27

thank you so much for watching truly do

28:30

appreciate you until next time

28:33

everybody bye

Rate This

5.0 / 5 (0 votes)

Verwandte Tags
Tesla FSDSelf-Driving CarsSan JoseVersion 12.3Traffic AnalysisParking Lot NavigationRoad ClosuresHuman-Like BehaviorAuto IndustryTech Review
Benötigen Sie eine Zusammenfassung auf Deutsch?