Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

FSD Beta 10.69

This site may earn commission on affiliate links.
Just finished my first drive with FSDb (10.69.3.1). Wow! Obviously needs to improve in a bunch of areas, but this technology is truly amazing! I had to intervene a couple of times, but that's what all those warnings were for when I accepted Beta. I'm happy to no longer have to tap the accelerator for every green light, the on-screen visualizations are incredible, and my car drove me all the way home with very little drama (even made it around a couple of roundabouts successfully).

I did discover that I prefer Chill mode, as "Normal" was a bit to aggressive for my taste. I think one key to successful widespread release will be to have the car drive a bit slower and smoother initially (to build confidence in the system). I think people will be patient if the car errors a bit on the side of caution, as long as it's smooth (like the way a Chauffeur would drive to minimize jerkiness).

It seems to me that many people on this forum get so deep into the weeds/details that they fail to recognize how truly transformational this technology is. It's definitely not 99.99% safer than an average driver (yet), but I can see a path to that level, and I'm happy to be an "early adopter" of this quickly developing technology!
 
For my initial neighborhood drive, 10.69.3.1 is improved but still rough.

Not enough regen use in city driving. Braking still feels like binary digital input instead of a smooth human analog. One case of excess brake and slow to initiate acceleration when lead vehicle slowed while changing lanes (release notes say it's improved). Acceleration also has programmed step-like feel which is mindless to implement but not ideal for comfort, confidence, or energy preservation.

My one lane selection edge case is slightly improved but still unsteady as following cars will question the slight left/right dithered steering while deciding path/lane choice. But at least it picked the correct lane.

The team needs to tighten the lateral G force comfort limit for turns. Rt turn still carries too much speed. In one case after the turn I experienced an abrupt brake (no vehicles or VRUs in sight) before accelerating - almost like an abrupt field of view change confused input image processing.

Unprotected right turn with heavy traffic still not able to make human-like decisive judgements. A trail of slowing oncoming cars with rt turn signals are assumed to continue through the intersection resulting in many missed merging opportunities.

One UPL did well minus the slow to jumpy step-like acceleration.

It's clear that driver comfort/G limiting/jerk limiting has not been much of a priority for Autopilot and FSD over the years. Jerky starts in stop and go are one big example, and that doesn't need sophisticated image recognition/detection. Compare that to competitors stop and go or adaptive cruise systems.

It's not that they don't know how how to fix jerky acceleration (straight line or after turns), it's that they haven't made it a priority. Stopping is another matter as it may be a detection of traffic thing, but even then the stops could be less sudden.
 
Right off the bat this morning, FSDb died as it approached the first stop sign. Red Hands of Death. It took about two minutes to recover.

After turning onto interstate frontage road, the two lanes merge to one. FSDb sort of dove into the other lane. Thought it was going to hit the guardrail. I let it go and all worked out, but it was dicey.
 
So it's only failing in situations that a driver unfamiliar with the intersection would also often fail at. I don't think any additional work on lane guidance will fix it, these particular intersections will likely need Tesla to implement fleet learning or memory of past traversals of the intersection in order to solve these.

That's the main deficiency: they need to make semantic maps for lane selection using human fleet driving. They have a potential asset far better than anyone else's but aren't using it.
 
  • Like
Reactions: willow_hiller
As others have said, I echo 10.69.3.1 is excessively cautious for street driving. I'd seen videos of excess caution for pedestrians and bicyclists but didn't expect it from other cars on the road.

One case of another vehicle slowing to a stop sign perpendicular to ego path triggered crazy heavy braking. I was traveling the 45mph speed limit when my costco items flew off the seats. It looked foolish. No human driver would interpret it as an impending accident. Thank goodness no one was closely following.

Unfortunately there's not much you can do to prepare yourself for this as it is totally unexpected.
 
  • Like
Reactions: kriskros91
It's clear that driver comfort/G limiting/jerk limiting has not been much of a priority for Autopilot and FSD over the years. Jerky starts in stop and go are one big example, and that doesn't need sophisticated image recognition/detection. Compare that to competitors stop and go or adaptive cruise systems.

It's not that they don't know how how to fix jerky acceleration (straight line or after turns), it's that they haven't made it a priority. Stopping is another matter as it may be a detection of traffic thing, but even then the stops could be less sudden.
No doubt. The last couple of releases they've been playing with jerk limits but more so for increased jerk. Ironically last year's AI day the planner team lead showed the planner 3 goals with comfort being one. One of these days they have to succumb but probably not before they figure out how to make the system more responsive without last second overcompensation of throttle, brake, and steering.
 
I am finding FSDb 10.69.3.1 to be too timid at stop sign intersections. I am usually intervening with the accelerator because the car with sits motionless at an empty intersection, or creeps at less than 10 mph through the intersection - with or without other visible traffic. This is a clear regression from 10.69.2.4.

I'm going to try switching to aggressive setting, though I don't expect an improvement.
 
Why does that always surprise you?
As described it was surprising. I have had many, many red wheels and disengagements and I’ve never received a strike.

It just seems difficult (at least in a car with functioning hardware, but that malfunction would be obvious).

Anyway, still waiting for those elusive good videos. Really curious how these things happen.
 
  • Like
Reactions: nvx1977
As long as I'm on a roll here. :)

That darn creep is just nuts. I've ever seen anyone drive like that other than Kramer's stop-short episode. Stop 30 ft before a stop sign? Really? I bet it's less safe and may even be illegal in some areas.

I still think the team is using it for extra processing time related to poor lateral oncoming vehicle path estimation which likely only worsens when ego abruptly decelerates to a stop.
 
Last edited:
  • Funny
Reactions: AlanSubie4Life
I know opinions of 10.69.3.1 are a dime-a-dozen these days, but I think it's still worth saying I'm relatively pleased with the improvements. Conservatively I'd say that it's about an 80% improvement in acceleration and braking in difficult spots (speed limit changes, cut offs, merging, last minute red lights), and a 50% improvement in lane selection. I still had one aggressive application of the brakes this morning, but it was at low speed after successfully routing around a landscaping truck that had stopped in a right turn lane. And I still had to disengage to manually put my car in the correct lane twice. But they're both situations with counterintuitive intersection design (one that randomly splits into an unmarked bus lane before rejoining the main road and immediately losing a lane, and one that requires the car to be in the far left turn lane, despite an immediate right afterwards).

So it's only failing in situations that a driver unfamiliar with the intersection would also often fail at. I don't think any additional work on lane guidance will fix it, these particular intersections will likely need Tesla to implement fleet learning or memory of past traversals of the intersection in order to solve these.
love the post 📫! Hardware 4 coming from TSMC, not Samsung.. production ramping in about 6 months... just in time for Cybertruck!
5/6 nanometers instead of 7 from Samsung. Coming from hw3 at 12 or 14 nn.
 
  • Like
Reactions: willow_hiller
On 69.3.1 Some improvement on slowing for oncoming traffic on winding roads. Smoother slowing for traffic controls. Still some very strange Phantom braking/ stopping. At pinned location on photo below always stops. I am unable to see any reason for stop rendered on screen. Lane selection is still a crap shoot. Today in far left lane sitting at a light with all three lanes full, turned on signal and tried t move to middle lane. Next turn was 7 mile ahead. From release notes for V11, it looks like lane selection should improve there.
 

Attachments

  • Phantom.jpg
    Phantom.jpg
    439.3 KB · Views: 54