An AP vs Nav possible bug?

[garygid]

Traveling under AP in the middle lane of 5 lanes, relatively light traffic,
good lane markings, no sun glare, the car suddenly wants to veer
sharply right into the empty lane. With my hands on the steering wheel,
holding very lightly, I detected the jerk in the steering wheel, grabbed
the wheel tightly, and corrected, probably before the car actually left
my lane. Fortunately there was no car to the right, but it could have been
serious.

No, I did not accidentally signal for a lane change, or use the other
controls, but having my fingers actually very lightly "gripping" the
wheel (to feel what AP was doing) allowed me react very quickly.

So, nothing visually apparent that would cause the AP to swerve
or jog right.

What I did do, a few seconds earlier, is give a voice command "Drive Home",
and I think the route was being drawn (approximately then) when this
unexpected "jog-right" event occured.

So, just maybe, something in route calculation or drawing, or calculation
of the expected energy usage graph, might have monopolized some
resource that caused the AP to "glitch"? Only a maybe, I realize.

In any case, a warning of an event with a high "pucker factor".

 

[BertL]

Interesting. Your description sure makes me think about challenges I (and many others) had many years ago with early virtual machine design on mainframes -- trying to figure out prioritization schemes for tasks needing common resources like compute power. I don't know enough how Tesla parses out compute power between the many multiple processors in MS, some of which are specialized, and how much is done in hardware vs in software, and then how Tesla tries to sync it all back up for UI, etc to speculate much further... but I've got to believe a lot of what Autopilot does is compute-intense, just as Nav is -- especially when a new best route has to be calculated... So there is an opportunity to run out of instantaneous compute power somewhere and priority has to be given to safety systems, including Autopilot when it's on for sure.

 

[CHG-ON]

I would definitely report it to Tesla. Thanks for the head's up.

 

[Patrick W]

Just a thought but if there had been a vehicle in the lane next to you I'm guessing AP would not have tried to move into that lane. At least I hope not.

 

[robby]

I had a similar "swerve" experience, but without having given any nav instruction. It's only happened once in about 450 miles of autopilot use. It occurred just after I passed a car to my right, and it was scary because the car was still in very close proximity to me when my car veered toward its lane. Like OP I was paying close attention and caught it very quickly, probably before I actually left my lane, but it's easy to imagine how one could be less lucky. I am going to pass this feedback along to Tesla later this weekend and offer to let them take the logs.

 

[Stoneymonster]

Interesting. Your description sure makes me think about challenges I (and many others) had many years ago with early virtual machine design on mainframes -- trying to figure out prioritization schemes for tasks needing common resources like compute power. I don't know enough how Tesla parses out compute power between the many multiple processors in MS, some of which are specialized, and how much is done in hardware vs in software, and then how Tesla tries to sync it all back up for UI, etc to speculate much further... but I've got to believe a lot of what Autopilot does is compute-intense, just as Nav is -- especially when a new best route has to be calculated... So there is an opportunity to run out of instantaneous compute power somewhere and priority has to be given to safety systems, including Autopilot when it's on for sure.

AP runs on customs CPUs (MIPS plus ASICs) from Mobileye. There is no contention.

 

[BertL]

AP runs on customs CPUs (MIPS plus ASICs) from Mobileye. There is no contention.

Perhaps, but there is more than AP being displayed on the IC for example, and some piece or pieces of hardware and more than likely software are coordinating things like those displays which will introduce points of intersection from a systems POV.

IMHO, only Tesla knows if something else under the covers may already be connected to AP, that isn't obvious to us as users quite yet. I used to do that all the time in my programming and systems days -- building interfaces between subsystems that were broader than the first set of end capabilities I delivered to my users, but took into account future design we had already solidified. It was a much more efficient way for us to move the ball forward. Isn't it fun when all of us, including myself, can only speculate on such things when it comes to what the Tesla Engineers are really doing under the covers with our MS?

 

[garygid]

Maybe some part of using the AP processed input requires a data link,
and the AP not available messages have to do with other parts of the
car using the data link, like Google maps and the traffic condition lines.

Just tracking the GPS Android smartphone locations will fairly quickly
draw a map. A new 270 degree on-ramp near me showed the traffic
movement path long before the Google map showed the new on-ramp.

 

[Stoneymonster]

The more likely cause is simply an AP vision or model error. It happens. I've had it fail to negotiate simple curves and decide to aim for a barrier, also with clear lane markings in perfect daylight. We must be cautious.