VLA 2.0 vs. FSD in Amsterdam – Part 2: Tesla Model 3

Support CleanTechnica’s work through a Substack subscription, on Patreon, or on Stripe. Help us produce all the original, high-quality content we publish week after week despite the challenges of AI deleting content, anti-social media, inflation and other obstacles.

In the first part of this series, we ride an XPENG L03 with VLA 2.0 in Amsterdam. For context, it’s best to read that article first. In the second part we move on to a Tesla Model 3 with FSD. While FSD is common in the US, it has not yet been approved in China, where XPENG’s VLA 2.0 is currently in circulation. The FSD is also not yet approved in most of Europe, but was first approved in the Netherlands and has been expanding to several other countries. As XPENG was testing VLA 2.0 in Amsterdam, this was the first opportunity to experience both systems on the same roads consecutively.

The car we were in was a black Model 3 Highland. This particular example lacked a turn signal lever. For Europe, the Model 3 is produced in Shanghai. The car was updated with the latest version of FSD, running on HW4. It was not a show-ready vehicle, but rather a customer car that the owner lent to XPENG. Tesla cars are very popular in the Netherlands, so we integrated easily.

Due to some difficulties with the camera, I ended up doing the test loop in the Tesla twice. This worked for the better. The same driver from video L03 was also able to join me on the second lap of the Tesla, sharing his perspective and providing consistency in the driver’s risk tolerance.

As with the XPENG trial, we faced challenges far beyond what most would experience in the U.S., from the proliferation of bicycles to complicated traffic patterns. However, since Tesla’s return was a little later in the day, traffic had slowed noticeably.

FSD: react and correct

With Tesla FSD, the general theme was react and then correct. If he saw a bicycle coming in the distance, he would react, brake sharply, and then move forward when he determined he had time to make a maneuver. It would tend to leave a slightly larger gap in the process. In some ways, this overly cautious approach could be considered safer.

However, those reactions could cause other problems. In one case, he stopped abruptly in the middle of the bike lane for an approaching cyclist some distance away and stayed there long enough to end up blocking a bike coming in the opposite direction on the other side of the bike lane. In another case, he stopped at a green light as a pedestrian approached an intersection but did not enter. By the time FSD realized what was happening, the light had turned from green to red. In another case, he stopped abruptly as someone approached an intersection while the light was changing, and then drove away when the light turned red. In one case, he was overly cautious on yellow and stopped when he could have crossed the intersection.

Of the two cars on the test track, only the Tesla had people honking. This was largely due to how it responded at intersections. Sometimes he was too cautious and held up traffic. Once, it stopped in the middle of the road and blocked the tram line and several lanes of traffic. At one waypoint during the first lap, it stopped in the middle of an intersection on a side street and did not reconnect.

The Tesla also had trouble choosing the right lane. On several occasions he turned into the wrong lane. In one case, he drove into the right turn lane as we were turning left and going around the block. The driver took over to return to the correct lane.

React and correct also applies to steering. In one case, on the first lap, he did not see a median early enough and had to change abruptly while turning. However, even in low-speed maneuvers, the steering wheel tends to move in spurts: turn, correct, turn, correct… It’s not as noticeable when the FSD works in isolation, but it appears as a significant difference after driving with VLA 2.0. When making some of those quick corrections, you could feel the tires slipping slightly on the wet, slippery bricks. I would like to test these systems again in snow or ice to see how they perform in even slipperier conditions.

“React and correct” has another double-edged nature. Harsh braking of a bicycle in the distance may be an error of caution for that bicycle, even if it goes further than necessary. However, the car behind you may not anticipate the sudden braking. As such, they may not stop in time and could rear-end you. This potential impact will likely be considered the other car’s fault, but a crash is still a crash.

When accelerating, the Tesla was not as smooth. In many cases, I would hesitate before taking off. Then it would initially go off the line and then slow down, then accelerate more. Overall, it was less comfortable. While it didn’t have a stopwatch, it probably didn’t accelerate faster overall to the desired speed. It just started later, took off faster, and then slowed down, rather than a constant rate of acceleration. Additionally, it would accelerate beyond the speed limit depending on the user’s settings, although that feature could disappear in Europe with the new UN DCAS regulation, which also requires hands on the wheel during urban driving.

As traffic thinned out as the day progressed, circumstances were a little different on my third loop. In areas with less traffic, the Tesla felt more comfortable. With less stopping and accelerating, the Tesla also felt smoother. Amsterdam is an extreme driving scene. Operating on more open roads with fewer obstacles, like most of the US, FSD would be a good choice. The latest version feels better than the version I drove a few months ago in the United States.

FSD had difficulty exiting and entering street parking spaces autonomously. In each case, the driver had to take charge. Part of this has to do with the tight parking spaces in Amsterdam. However, upon attempting to self-park, FSD indicated that it had successfully parked, but was still in traffic. The driver had to park closer to the sidewalk again. This also might not be a big concern in the United States, where spaces are not as tight.

The Tesla was able to clear obstacles with a gap of one or two inches. However, it was not a quick or smooth process, as multiple corrections were made while progressing slowly. If given enough time, he could perform the maneuver. That said, it would be a challenge if the car got halfway and then gave up, as it might be difficult for the human to get out of the situation. Due to the design of the vehicle, it is a little more difficult to know where the curves are, especially when the driver is less familiar with the car. Luckily, we didn’t get caught in any bind.

Waypoints were a challenge for the Tesla. Sometimes I would lose my reference point and then try to turn around. In one case, he missed the waypoint, turned around, missed it again, and then tried to take another road to make another loop. In that case, the driver took control and then re-engaged the FSD when the waypoint was cleared and the car was moving in the correct direction again. In another, he attempted to stop at a landmark in the middle of a road, blocking traffic. Then when the waypoint was deleted, sometimes the FSD would not reconnect or take a few seconds to take off again. In general, many drivers do not use reference points while driving. However, as vehicles approach higher levels of autonomy, the ability to navigate to specific points becomes more important.

When maneuvering through construction-related obstacles, the Tesla would sometimes make a wrong move, resulting in a takeover to get back on the right path. We also stayed in the wrong lane for a considerable amount of time after climbing a curb to avoid a stopped vehicle. The Tesla waited for another curb stop to return to the correct side of the road. We passed a police officer coming toward us in our lane, but fortunately he didn’t turn around to give us a ticket.

However, there was also an instance where the Tesla backed up to make room for approaching construction equipment. That was impressive. Reversing will be necessary on some narrow single-lane roads in Europe, as seen in some parts of Greece. Reverse operation hasn’t been enabled on the L03 prototype yet, but that gives XPENG something to work on.

Overall, the Tesla felt like a middle-American driver trying to drive in Amsterdam. The traffic patterns are more complex than New York, and much more complex than a city like Austin. He felt overly cautious and nervous, not being fully adapted to the complex traffic and road conditions. In different circumstances and with a different frame of reference, you would probably feel more comfortable.

In part three of this series, I’ll go into more detail about the cars and compare some of the differences. But you can also compare it with the previous XPENG L03 video: