Vermont to Florida -- the slow way -- on FSD 10.3

[BizJet]

A long-planned trip in my Model Y was due to start at 7:30am Sunday. I'd carefully guarded a 99 Safety Score over 1200 miles in hopes that I'd get the FSD beta and be able to write about a nearly 2,000 mile drive from Quechee, Vermont, to Naples, Florida on FSD 10.3. A 100 score on my Model 3 had qualified me for FSD 10.2 a couple of weeks ago, and I'd logged a few hundred miles on it, including some rather interesting miles through Boston traffic. I had high hopes for FSD 10.3 on my Model Y, and wanted to write about the trip, in which I expected to encounter a wide variety of conditions, from Vermont gravel roads to Virginia's Blue Ridge Parkway to the Gillette Stadium parking lot to Atlanta city traffic.

Two disappointing delays later, lo and behold FSD 10.3 was miraculously available exactly 4 hours before departure, and I was ready to go. In this thread, I'll chronicle my observations through this tremendous range of driving experiences on an 1800-mile trip on FSD.

Leg 1 -- Sunday morning, October 24 -- Quechee, Vermont to Foxborough, Massachusetts (171 miles)

First things first on the trip. I had tickets to the New England Patriots game against the New York Jets, so the first stop was at Gillette Stadium in Foxborough, Massachusetts.

• FSD 10.3 did a decent job on the gravel back road that leads from Quechee to Woodstock, VT. Unlike its 10.2 predecessor, it stuck to the right side of the road, and even occasionally displayed an imaginary center-stripe on the road, staying to the right of it. (Previously, 10.2 would drive down the center of the road, creating some harrowing moments around blind curves with oncoming cars. There is a possibility that this was not a software upgrade, but as result of the fact that there were fewer fall leaves on the roadway this time I drove it, owing to the fact that fall is almost done in Vermont.)
• I was wondering how FSD would handle one of Vermont's famous one-lane covered bridges. It did fine, though an interesting edge case for the future will be how it handles the case where cars take turns on either end of a long bridge.
• There was considerable fog in the Connecticut River Valley, which limited visibility. Autopilot decreased speed to 45 mph on a 65 mph road, which felt unnecessary given how far I could see ahead. I took over and drove at 65mph for a few miles.
• South on I-91, I couldn't figure out why my car almost blew past the programmed exit. Then it occurred to me: interstate highway driving is done on an entirely different autopilot stack -- basically the old autopilot we've been using for a couple of years. It's necessary to press "Navigate on Autopilot" (or enable it by default in settings) in order for autopilot to recognize exits. I'd already been spoiled by my experience with FSD 10.2, which requires no such action on side roads.
• On state highways, FSD does NOT see or comply with yellow speed limit signs, such as those marking a curved section of road, a construction zone, industrial zones, etc. I also do not believe it recognizes active school zones, which are strictly enforced in New England at 20mph.
• When transitioning from a high-speed area to lower speed zone, such as coming into a town along a highway, the car does not BEGIN to slow down until it has passed the new speed limit sign. And even then, the deceleration is quite gradual. On several occasions, the speed limit went from 55 to 30, and the car was still going 45 mph 200 yards past the speed limit change. In each of these cases, the speed limit change was foreshadowed by a "Speed 30 ahead" sign 1/4 mile before the actual speed change; a human driver would have started slowing down well ahead of the speed limit change.
• FSD slowed down a whole lot -- it possibly even would have stopped -- before entering two narrow rotaries in Keane, New Hampshire, even though there was no threatening traffic in the rotary. Tapping the accelerator seemed to give autopilot more confidence to proceed, and it handle the rotary just fine.
• Speaking of rotaries, FSD had more trouble with large, wide-open roundabouts that have a park in the middle, such as is typical in rural New England. In particular, the different lane configurations for exiting or continuing in the roundabout were confusing.
• Once I reached the stadium, FSD seemed to see all of the pedestrians walking along the road, though it wanted to drive uncomfortably close to them when they were walking partially in the road.

That said, I was impressed. I saw none of the faux Foward Collision Warnings or Panic Braking that had been reported by others on 10.3, and FSD handled most situations comfortably -- even better than 10.2.

Bad News at the Stadium - FSD 10.3 goes away!

Good news/bad news at Foxboro. The good was that the Patriots trounced the Jets. The bad news is that I saw during the game that a new software update was available, and, thinking that it was further improvement on 10.3, I accepted the update and -- POOF -- FSD 10.3 was gone, having been recalled.

So I've driven through Rhode Island, Connecticut, New York, New Jersey, and Pennsylvania on the old autopilot system, and as I approached Virginia's Shenandoah Valley, I've just gotten notification that FSD 10.3.1 is backed -- presumably fixed. I'm installing it now, and that will take me to the next installment of my observations as I enter Skyline Drive and the Blue Ridge Parkway, which I'll post tonight or tomorrow!

 

[BizJet]

Leg 2 -- Sunday evening, October 24 -- Foxborough, Massachusetts to Bernardsville, New Jersey - 251 miles (422 so far)

Because of the temporary recall of FSD 10.3, this drive was done on the old highway FSD stack, though it was 99.9% autopilot nonetheless.

• One thing that was noteworthy, though, is that even though it was on autopilot, the Safety Score got clobbered (84) during the second half of this drive, much of which was in stop-and-go traffic on I-95 in Southern Connecticut. It appears that while the Safety Score masks forward collision warnings incurred while on autopilot, it does pick up hard braking, some of which happened on autopilot during the long stop-and-go period (6.1% of the braking was considered "hard", which accounts for the 84% safety score on the drive).
• Bernards Inn has free Tesla destination charging for overnight guests.

Leg 3 -- Monday morning, October 25 -- Bernardsville, New Jersey to Woodstock, Virginia - 287 miles (709 so far)

My longest drive ever in a Tesla. Left New Jersey with 100% and arrived at the Woodstock Supercharger with 1 mile left in the tank (carefully planned). Again, because of the FSD 10.3 recall, this leg was also done on the previous autopilot. No big deal, because again this was all interstate highway driving and would have been done on the highway autopilot stack anyway. The interesting part was going to be the next section, which would be on winding, scenic mountain roads.

Leg 4 -- Monday mid-day, October 25 -- Woodstock, Virginia to Elkton, Virginia - 26 miles (735 so far)

Right on cue, the notification came that a new version was downloading -- slowly, because much of the Shenandoah Valley has only 3G cell coverage. I continued south as it downloaded, and by the time I hit Elkton, the software was ready to install and I was ready for a break.

Leg 5 -- Monday afternoon, October 25 -- Elkton, Virginia to Vesuvius, Virginia -- 103 miles (838 so far)

I wasn't going to give FSD 10.3 a chance to wait. I would throw at it the challenge of driving one of the most scenic, winding mountain drives in America -- Skyline Drive in Shenandoah National Park, and the northern section of the Blue Ridge Parkway. The results were mixed.

• On the one hand, the visualizations seemed to be highly accurate. In particular Skyline Drive is well-marked with a yellow double center stripe and white edging, and Tesla Vision seemed to see absolutely everything. The Blue Ridge Parkway has the yellow centerline but does not mark the edge of the road except at the scenic overlooks, but the visualizations looked rock-solid. Following the visualizations, however, was a different matter.
• First of all, FSD seems to have a lot of trouble with the combination of bright sunshine and dappled shadows of tree branches on the ground. The car seemed to break frequently when there was rapid transition from sun to shadow, as if it wasn't certain whether it was seeing shadow or an object in the road. This made for a rather jerky, not-too-comfortable-and-smooth ride.
• The car slows down painfully around sharp right-hand bends in the road, as if it expects to suddenly encounter some obstacle around the bend. Corners designed to be taken at 45 mph were navigated at 25 mph. Yes, the caution was admirable; but a human driver would be closer to the 45, because the corners weren't THAT blind.
• Several times, the car would wander randomly into the left (oncoming) lane on a right bend, requiring correction. No particular reason was observed; these clips have been captured and submitted.
• The net effect was that a road that was designed for 35 mph was navigated at an average of about 25 mph, and the parkway whose speed limit is 45 was traversed at about 35 mph because of all the slowdowns.
• At the end of my day, I inched down a very narrow gravel road to the Sugar Tree Inn, chosen because it has a Tesla Destination Charger for guests. The road is posted (by a non-standard sign that the car didn't pick up) at 20mph, but 15mph was actually a more appropriate speed. For whatever reason, the car set the speed limit at 50 mph. I had to disengage several times down the steep, winding hill, and each time I re-engaged the set speed shot up to 50mph, and I had to quickly dial it back down to 15mph. It feels like if a speed is overridden and autopilot is briefly disengaged, the set speed should resume where it was before, and not reset to the (incorrect) speed limit displayed.

As I re-read this, it appears that I'm down on FSD 10.3, in that I've focused on the problems I encountered. Quite the contrary -- I am an enthusiast and continue to be impressed with how FSD performs 95% of the time. Yesterday and today, I threw a whole lot of difficult problems and edge cases at it, and it performed probably the same as one of my teenagers would have if I had taken them on the same roads with a fresh learners permit. I am totally dedicated to doing my part to chronicle and report the issues I encounter and to continue to pose interesting cases that can serve as learning experiences for the neural network.

Kudos to the engineers at Tesla for getting us this far. And if you haven't taken the time to watch the Tesla AI day video (

) to see what has gone into FSD, I recommend that you do so -- it really helped me understand what might be going on when I saw FSD doing something unexpected.

Tomorrow, a full day of Blue Ridge Parkway headed toward Asheville, North Carolina.

 

[EVRider-FL]

I drove on Skyline Drive and Blue Ridge Parkway a couple of years ago with Autosteer, and I didn’t experience some of the issues you did with FSD Beta. The car did slow down too much on some curves, but overall it did surprisingly well. I had to disable Autosteer at times on the Blue Ridge Parkway because traffic was moving faster than my maximum of 5mph over the speed limit.

 

[Jejunjm]

Thanks for the write up. Much more enjoyable for me then any YouTube video.