Hands-On With XPeng City NGP Self-Driving

What is it?

We first drove the XPeng P5 back in fall 2021, but we weren’t able to take full advantage of one of its biggest selling points; LIDAR. At the time, the company’s LIDAR-powered driver assistance system, then called NGP 3.5, was still under development. Nearly a year later, XPeng invited us to Guangzhou so we could finally get behind the wheel of a P5 that took advantage of LIDAR to achieve the company’s highest level of autonomy yet, now called City Navigation Guided Pilot (CNGP). We were put behind the wheel of the as system as it drove us around the busy streets of Guangzhou, a city of 15 million people.

Sensor fusion

The XPeng P5 has a suite of 32 perception sensors, including 13 high-definition cameras, 12 ultrasonic radars, 5 millimeter wave radars, and 1 submeter high precision positioning unit. However, the stars of the show are undoubtedly the twin double prism LIDAR units that are mounted on either side of the front bumper.

According to XPeng, by fusing LIDAR and visual sensors, they're able to achieve higher levels of accuracy in recognizing obstacles, pedestrians, and vehicles on the roadway. That’s especially true at night, in tunnels, and in snowy, foggy, and rainy conditions. XPeng was also quick to point out that they are currently using 20 to 30 TOPS computing power to achieve this level of autonomy.

Familiar and not so familiar

Speaking of conditions, we arrived in Guangzhou on the winds of a typhoon, which was supposed to bring heavy rain the next day. Thankfully for XPeng, we only saw some mild showers. During our experience, we were accompanied by an XPeng safety engineer, who was there to answer questions and, as it turns out, tell us when to intervene with the system.

The process for activating CNGP is the same as pretty much any other navigation on autopilot system: set a destination in the navigation, and once the system prompts you that CNGP is available, pull down on the shift lever twice. We set out from XPeng headquarters and were almost immediately prompted by the car to activate the system.

Seconds after we activated CNGP, the car made a lane change to prepare for a U-turn up ahead. We’ve experimented with a variety of driver assistance systems, including Tesla FSD, GM Super Cruise, BYD DiPilot, and others, but this was the first time we’ve ever used a system that was capable of driving in these congested, urban conditions. The sensation was quite strange, at least at first.

Using CNGP on the busy streets of Guangzhou was highly illustrative of something Chinese manufacturers point out all the time; the relative complexity of China’s roads. Roadways everywhere are complex, but China’s mix of passenger vehicles, buses, pedestrians, electric scooters, and bicycles means it can be more challenging than places like the United States or Europe.

It wasn’t long before the system was put to it’s first test, stoplight recognition. The P5 uses its high-definition maps to know in advance where stoplights will be located, and then “reads” those stoplights using its 8MP cameras. We were admittedly nervous as we approached our first light, but the system brought us safely to a stop and accelerated away from the light with no problem. That would continue to be the case throughout the trip, though we did find that it tended to continue accelerating through left hand lights even if the light changed to yellow well before it had exited the turn lane. That’s what most human drivers would do, but it felt a little bit to bold for a system that’s this early in its development.

Merging from one roadway to another is something that other NOA systems have been capable of for some time, so it’s no surprise that the P5 accomplished that task several times during our drive. The system was clearly designed to be conservative when changing lanes, but it was still able to merge each time, and without holding up everyone else on the road.

The turning inputs, whether it was making a lane change, doing a U-turn, or turning at an intersection, were precise each time, with no sense of hesitancy. Nor did it make any of those annoying micro adjustments that some systems seem to make mid-turn.

Acceleration was equally impressive, with the car getting up to speed at a rate that felt in line with a human driver. It tended to stay at or below the speed limit, which sometimes made it hard to keep up with traffic. Having said that, we would hope that engineers wouldn’t be programming the vehicle to do it any other way.

There were multiple instances in which the car was forced to brake rather suddenly due to other drivers who wanted to merge late, but it never felt as though the system was overreacting. Having said that, brake modulation was the weakest link in the entire experience. The car didn’t fail to brake at any point, but it did tend to brake a harshly in certain situations, such as when approaching a stoplight.

The dreaded roundabout

One of the more stimulating moments of the trip, and the first instance in which we intervened, occurred about 20 minutes into the drive. We had already navigated a variety of environments, from narrow U-turns to busy four-lane roads, when we approached a large roundabout below an expressway. This was a very complicated driving situation, with cars constantly exiting and merging onto the roundabout. The car was navigating without incident, until we came upon a collision between a large cement truck and a Roewe sedan that had occurred in the first and second lanes.

At first, we planned to let the system try and figure it out on its own. However, the XPeng safety engineer in the back seat, who presumably has a great deal of experience with the system, suggested that we take over instead. He explained that he wasn’t sure how long the system would take to comprehend and overcome this obstacle, and that could leave us sitting behind it for quite some time.

We had a similar experience earlier in the day when were given a ride in the P5 with a different engineer. While crossing a divided two-lane bridge, we came upon two work trucks that had stopped in the righthand lane, blocking our progress. The car didn’t have time to change lanes, as there was traffic in the left lane, so it stopped behind the trucks and waited for an opportunity to pass.

The amount of traffic approaching from the rear was such that it took us 2-3 minutes to get around, as the car made several aborted attempts to change lanes, but then detected approaching traffic. A human driver would have been able to do it much faster by inching into the left lane until an approaching car came to a stop. Not an instance in which we felt unsafe, but a sign that this system isn’t going to get you where you want to go as fast as you could do it on your own.

Beware student drivers

We were among the first people outside of XPeng to be able to experience this system at night, when it would rely more heavily on its LIDAR, instead of its cameras. LIDAR is not affected by lowlight situations, but can be affected by things like rain. As it happens, it was raining during our drive, further increasing the challenge to the system.

To its credit, CNGP’s performance after dark was indistinguishable from our experience during the day. There were no noticeable differences when it came to detecting stoplights, people, or objects on the road. We did have to intervene during our 30-minute drive, but the circumstances were rather unique.

We were headed down a busy road in heavy traffic when we began to approach a group of student drivers from behind. The group had formed a sort of moving roadblock, but other drivers were managing to find ways around them. Our car, however, found itself stuck behind the one of them in the middle lane traveling well below the speed limit for several minutes as the system failed to make a lane change and go around. We felt that there was plenty between it and approaching traffic, but it clearly disagreed.

Finally, our car began to change to the left lane to overtake the car ahead, just as that car decided to make a lane change as well. At this, point, we realized there was an even larger group of student drivers in that left lane, so we intervened to cancel the lane change and continue in the middle lane. It’s hard to blame the system entirely in this situation, as there is little reason to expect it to recognize student drivers on the road. Having said that, we would expect it to be slightly more aggressive in the way it changes lanes the longer it has to wait behind another car.

Takeaway

After spending around two hours driving (riding in?) the XPeng P5 equipped with CNGP, we were left very impressed by a vast majority of the experience. Does that mean the system is ready for primetime? No. While its performance was impressive, it still feels like a work in progress. When it comes to systems that drivers need to rely for their safety and the safety of everyone else on the road, that’s not quite good enough.

XPeng seems to think it’s ready for a rollout, albeit a very small one. They have announced that they will be releasing CNGP to a select group of P5 owners in Guangzhou with plans to make it available in other major cities in the future. No details were provided regarding how the owners would be chosen, where within Guangzhou they would be able to utilize the system, or how long the rollout to other cities would take.