-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Autopilot is already improving.
- Thread starter mobe
- Start date
Vger
Active Member
My wife and I just completed a 450 mile drive with AP, a route we have done many times before, and know almost mile by mile. This route extends from south of Vancouver, BC (BC rt. 99, US I-5) through Bellevue, WA (I-405), over Snoqualmie Pass to Vantage, WA (I-90), east on WA-26, to WA-17, WA-260, WA-261 and finally US-12 to Clarkston, WA, our destination. The first 275 miles of the trip are on divided, limited-access highways; the last 175 miles are on well-maintained, well-marked two-lane highways.
We were able to maintain AP function on at least 95% of this route, both on the divided 4+ lane segments and undivided 2-lane segments. Overall, the AP functioned incredibly well. I was very alert throughout and kept my hands on or touching the wheel throughout, as Tesla instructs. In fact, I got quite a few "hold the wheel" reminders. I noted that the frequency of the reminders seems to depend on the "stress" that the AP is currently experiencing, which relates, I believe, to both the clarity of lane marks, but also to the level and speed of traffic. When on the two-lane highways, I was particularly careful, especially when approaching oncoming traffic. At those times, I had my hands firmly on the upper handholds of the wheel, ready to instantly counteract any deviation by AP toward the centreline. No such issues occurred.
On the divided highways, we passed dozens of exit ramps while in the right lane. Most were passed with no drama at all. In Canada, exits ALWAYS have a dotted line across them, so are never an issue. In the U.S., most exits do NOT have the dotted line. On a number of these, AP made a slight deviation (<10 cm) toward the exit, but immediately corrected itself before I might have taken over. In only a two or three cases did I feel that AP might take the exit, and I took control. So exit ramp performance was better than 95%. This was our first AP trip on this route, but these highways are travelled fairly heavily by other Teslas. So I believe we may have been the beneficiaries of "fleet learning" and especially fleet-mapping of these segments and their exits.
My experience of auto-lane-change was fabulous. I had my doubts about the feature going into the drive, but it quickly became one of my favourite AP capabilities. Transitions were always very smooth and stable, even when executed on the sweeping curves and grades of Snoqualmie Pass.
Now for the two-lane highways: AP performance overall was almost as smooth, steady and reliable as on the divided highways. We were both amazed to find the the AP could smoothly and safely cross many 90-degree intersections where the lane marking disappeared for the 20-30 metres of the intersection. Occasionally AP would wobble very slightly, but there were never any lurches or panics through these intersections (no stop signs or lights).
WA-26 has a number of passing lane sections where the two lane highway adds a temporary third lane. When we approached these sections, I found that very gentle pressure on the wheel served to bias AP into the right lane, and out of the passing lane. The AP would hunt and waver for a moment before this as the center and edge lane marks diverged to create the new lane. At the end of these passing lane sections, when the two lanes became one, AP would wobble again momentarily when the dotted line disappeared, and before it picked up the center line on the left and the edge line on the right, but quick it would lock in again, and proceed smoothly.
In 450 miles, we only disengaged AP for the 25 miles of very twisty WA-261, which is riddled with tight curves marked down to 40, 30, even 20 mph. The rest of the time AP ran for hours without intervention. Of course, I took over on exit ramps and through intersection turns and merges. There were only two or three red-warning "take control immediately" events, and those were clearly very ambiguous intersections or highway transitions. None were surprising or unexpected, and I always had my hands resting on the wheel when needed.
When one uses AutoPilot as intended and instructed, it is amazingly reliable already, and hence one can relax. While avoiding extreme tests and stunts, I can well appreciate Tesla's accomplishment, and their confidence in turning us loose on it.
To top off this positive experience, today we took my wife's 89-year-old Dad for a drive on US-12. He loves our Teslas and always looks forward to riding in them. I told him that the car had just learned some new tricks, and then engaged AP and took my hands off the wheel to make sure he realized what was happening. After the first sweeping curve at 60 mph, he looked at me dumbfounded, and said, "well I'll be damned!" (He is a U.S. Navy veteran, hence the salt.)
'Nuf said.
- - - Updated - - -
Cool, we saw the same thing here in Clarkston, WA.
Machine learning and crowd-mapping; I tell you, it is alive!
We were able to maintain AP function on at least 95% of this route, both on the divided 4+ lane segments and undivided 2-lane segments. Overall, the AP functioned incredibly well. I was very alert throughout and kept my hands on or touching the wheel throughout, as Tesla instructs. In fact, I got quite a few "hold the wheel" reminders. I noted that the frequency of the reminders seems to depend on the "stress" that the AP is currently experiencing, which relates, I believe, to both the clarity of lane marks, but also to the level and speed of traffic. When on the two-lane highways, I was particularly careful, especially when approaching oncoming traffic. At those times, I had my hands firmly on the upper handholds of the wheel, ready to instantly counteract any deviation by AP toward the centreline. No such issues occurred.
On the divided highways, we passed dozens of exit ramps while in the right lane. Most were passed with no drama at all. In Canada, exits ALWAYS have a dotted line across them, so are never an issue. In the U.S., most exits do NOT have the dotted line. On a number of these, AP made a slight deviation (<10 cm) toward the exit, but immediately corrected itself before I might have taken over. In only a two or three cases did I feel that AP might take the exit, and I took control. So exit ramp performance was better than 95%. This was our first AP trip on this route, but these highways are travelled fairly heavily by other Teslas. So I believe we may have been the beneficiaries of "fleet learning" and especially fleet-mapping of these segments and their exits.
My experience of auto-lane-change was fabulous. I had my doubts about the feature going into the drive, but it quickly became one of my favourite AP capabilities. Transitions were always very smooth and stable, even when executed on the sweeping curves and grades of Snoqualmie Pass.
Now for the two-lane highways: AP performance overall was almost as smooth, steady and reliable as on the divided highways. We were both amazed to find the the AP could smoothly and safely cross many 90-degree intersections where the lane marking disappeared for the 20-30 metres of the intersection. Occasionally AP would wobble very slightly, but there were never any lurches or panics through these intersections (no stop signs or lights).
WA-26 has a number of passing lane sections where the two lane highway adds a temporary third lane. When we approached these sections, I found that very gentle pressure on the wheel served to bias AP into the right lane, and out of the passing lane. The AP would hunt and waver for a moment before this as the center and edge lane marks diverged to create the new lane. At the end of these passing lane sections, when the two lanes became one, AP would wobble again momentarily when the dotted line disappeared, and before it picked up the center line on the left and the edge line on the right, but quick it would lock in again, and proceed smoothly.
In 450 miles, we only disengaged AP for the 25 miles of very twisty WA-261, which is riddled with tight curves marked down to 40, 30, even 20 mph. The rest of the time AP ran for hours without intervention. Of course, I took over on exit ramps and through intersection turns and merges. There were only two or three red-warning "take control immediately" events, and those were clearly very ambiguous intersections or highway transitions. None were surprising or unexpected, and I always had my hands resting on the wheel when needed.
When one uses AutoPilot as intended and instructed, it is amazingly reliable already, and hence one can relax. While avoiding extreme tests and stunts, I can well appreciate Tesla's accomplishment, and their confidence in turning us loose on it.
To top off this positive experience, today we took my wife's 89-year-old Dad for a drive on US-12. He loves our Teslas and always looks forward to riding in them. I told him that the car had just learned some new tricks, and then engaged AP and took my hands off the wheel to make sure he realized what was happening. After the first sweeping curve at 60 mph, he looked at me dumbfounded, and said, "well I'll be damned!" (He is a U.S. Navy veteran, hence the salt.)
'Nuf said.
- - - Updated - - -
Cool, we saw the same thing here in Clarkston, WA.
Machine learning and crowd-mapping; I tell you, it is alive!
Last edited:
Vger, thanks for the thorough and informative report. What a great experience it must have been.
What were your thoughts on how you felt from a fatigue level after using AP on this trip? I'm assuming it was relatively higher stress than it'll be once your confidence in it goes up, but did you notice any difference in how you felt vs. having to manage the vehicle position in the lane?
Vger
Active Member
My recollection of your neighbourhood, Steve, (and a happy one) is that there are quite a few tight-ish curves. If I recall correctly, that might have something to do with it.
- - - Updated - - -
Great question, and I have been pondering it already. We got up at 5am to catch the ferry for the drive, so I was a little groggy from that, and just getting over a cold. So my energy level was not fabulous to begin with, and I was tired after 12+ hours at the end.
BUT, what I experienced is that once I gained confidence, I did relax, especially physically at the shoulders. I also sense that when my attention and brain resources are not going into guiding the car on a second-by-second basis, I am more free to truly supervise the drive, and devote more energy to analyzing threats in the traffic and environment. This could be one of the keys to the safety benefits of AP. Not just that it will intervene for you, but that you have more bandwidth to intervene when you are less occupied with routine control.
There is another tesla owner up here as well. It will be very interesting to see as you get more passes if it eventually starts working.
I find an interesting situation for the car is when the road has a curve on a crest. The crest may only be 4 feet high but it can't see the curve beyond the crest, it can see a parked car or even a oncoming car beyond and think that is where the road heads so will direct it self towards this collision. I have a few points on my daily run where this is the case. I keep my hands on the wheel and I look forward to the car learning. Has anyone else had the same experience?
BertL
Active Member
Yes, similar thing happened to me yesterday on a dual-lane freeway on-ramp that curved to the left while gaining altitude. As the vehicle in my right lane sped-up slightly ahead of me, all of a sudden my MS slowed down as if it was going to strike that car in the adjoining lane, but as soon as the curve was over and that car had pulled-away, Autosteer picked-up speed on it's own again.
Vger
Active Member
Yes, absolutely. This is a clear weakness of the current system. It MAY be aided by more high-precision mapping, but I wonder if the field of view of the sensor suite is fundamentally limiting.
During our long drive chronicled upthread, this was the biggest limitation on WA-261. Rather than engaging AP and nervously watching it repeatedly fail, we simple watched the display to see when it lost lock on or recognition of the lane edges. On essentially every crest (there were dozens on this road), the display showed that the car had no clue where the road was going.
This is a situation where (alert and well-trained) humans are really good at predicting the path of the road without direct vision of it. It will be interesting to see if the AP can be taught to do this, or if it will depend on the explicit detailed mapping of such roads.
If I could provide some none-Tesla context for autopilot...
As previously mentioned, wife has a 2015 CRV Touring. It has Honda's current best lane keeping, TACC and emergency braking etc.
Experiences so far sound similar to autopilot in many ways, except the crucial one.
When an exit ramp appears it often tries to exit and needs encouragement to stay in lane.
Sharp curves fool it.
Sometimes it loses track of the car ahead and either gives up to speeds up to match with a new car.
All of which are similar to AP - except it almost feels like driving using an xbox controller
The biggie though is that my wife's CRV won't get any better unless she sells it and gets a newer car.
As previously mentioned, wife has a 2015 CRV Touring. It has Honda's current best lane keeping, TACC and emergency braking etc.
Experiences so far sound similar to autopilot in many ways, except the crucial one.
When an exit ramp appears it often tries to exit and needs encouragement to stay in lane.
Sharp curves fool it.
Sometimes it loses track of the car ahead and either gives up to speeds up to match with a new car.
All of which are similar to AP - except it almost feels like driving using an xbox controller
The biggie though is that my wife's CRV won't get any better unless she sells it and gets a newer car.
Andyw2100
Well-Known Member
I have a similar experience.
There is one curve in my regular drive that I pay special attention to. It is a connection from one highway to another. When I first took delivery of my P85D, I worked up to learning that the car could easily handle this 45 MPH suggested speed curve at 70 MPH--the cruise speed I would typically have set on this 65 MPH highway. For months now I've been taking this curve at 70 with TACC engaged whenever I can, which means whenever there isn't a vehicle in front of me that will take it slower and slow me down. The curve has some elevation to it at the beginning, and then slopes down to meet the other highway.
The first time I tried to take this curve with TACC and Auto Steer Beta engaged was on 10/16. I was not surprised at all when very early in the curve the car threw up the "Take Control Immediately" alert.
I tried again on 10/24, but I had slower traffic in front of me, and while the car handled the curve just fine without any intervention from me, I think we took it at about 50 MPH, so I wasn't particularly impressed, and didn't think anything of it.
But today was different, and impressive. I had no traffic in front of me, and approached the curve with TACC set to 70 MPH, and Auto Steer Beta engaged, with my hands very lightly on the wheel, fully expecting the full abort alert. I was shocked to see the car begin to slow just a little, and part way into the curve I received the mellow request to put my hands on the wheel, so I lightly gripped the wheel, and the request went away before we were through the curve. The car never slowed below 67, and except for asking me to place my hands on the wheel briefly, did almost everything on its own. This was a marked improvement from the first time, when Auto Steer Beta completely aborted.
I'll be interested to see if over time the alert to put my hands on the wheel eventually isn't necessary, and/or if the set TACC speed of 70 can be maintained. But it is clear to me that somehow, some way the Autopilot system is improving.
sillydriver
Member
All of the reports in this thread are great news. I just wish we had some indication of when and how (new build or whatever) to expect Autopilot v1.01! Or have we somehow already got it?
Cyclone
Cyclonic Member ((.oO))
If things are regularly improving, why would you car? It'll work or it won't either way.
I don't think you can associate all changes with updated firmware. Supposedly the AP models will be updated constantly with new data. There's no need to for them to upgrade the entire firmware or release new firmware builds when just data is changed.
sillydriver
Member
:biggrin: If I knew it was just installed, then I would have an excuse to go out and flog the system, trying all the situations I know are marginal, just like a proper (if amateur) beta tester.
Cyclone
Cyclonic Member ((.oO))
But you do know that the fleet learns from each other and there are regular updates that don't require firmware updates, right? It could be improving daily or weekly without an update, so feel free to drive and drive every day!
I continue to wonder how that's being pushed to the cars. Elon's tweet about Autopilot 1.01 almost makes it sound like that is the new model. But the reports here are pretty strongly making the case that things are already updating. I wish I had an AP car so I could set up a full-time packet capture and figure out what's going on.
If we assume that indeed models are being pushed out daily to the vehicles, then the AP 1.01 could really just be the interpretive interface between the model and the vehicle. It could provide better cleaned data back to HQ, and it could have the vehicle respond differently to the signals from the model. If so, I imagine it'll be bundled with a firmware release.
From what I've been reading here, I think each car is just learning from its own AP driving experiences and driver feedback. People who drive the same route day after day report improvements; but I just test drove an MS that jerked left on Highway 1 for no reason that I could see - the sort of thing that people report happening less and less when they go over the same route. If all the MS's that have driven up and down that section of Highway 1 were putting their learnings together and sending it to this car, I wouldn't have expected that jerk. But this demo may have never done this section of road with AP, so it behaved like people reported with their initial AP performance.
Likewise, a server taking the learnings of all the AP cars out there, processing them to improve guidance/algorithms, and then sending that new info back to every MS on a ~daily basis (without seeming to transmit a lot of data) seems like a lot to expect from a system that's just rolled out.
So my guess is that it's "local learning" that's generating the rapid AP improvement many are reporting, and AP 1.01 will be the first time since AP 1.0 that individual MSs receive "crowdsourced" data or improved algorithms from other MSs.
Vger
Active Member
Respectfully, the problem with this interpretation is that it directly contradicts what Elon said when he introduced the system to the press. If you have not listened to the press conference, you must do it to understand this thread.
Many of us have seen evidence of improved performance with such things as exit ramp avoidance on roads that OUR car has never driven before. I believe that fleet learning, as described by Elon, is real.
I know this is all speculation but ^^ this makes the most sense to me. If the learned data was localized to the car it would require processing and storage which would be unnecessary and may bog things down.
Last edited:
Similar threads
