What are the shortcomings of Waymo's self-driving cars?

Recently, The Information released a report pointing out that Waymo's self-driving car road tests are causing a lot of trouble for local people, and that some driving habits are very different from those of human drivers. Even the fastest-growing Google self-driving car cannot currently be well integrated into daily traffic.

Waymo is the leader in the self-driving car race, with its self-driving cars having driven a total of 5 million kilometers in the real world and 5 billion kilometers in simulation as of May this year. However, a recent report from The InformaTIon pointed out that Waymo self-driving cars are currently experiencing some difficulties in Chandler, Arizona. The report shows that even Google's self-driving cars still have many problems when encountering other road users.

The report claims that the most obvious problem of Waymo's driverless cars is turning left. The report states: "Waymo's driverless cars often have trouble in busy traffic areas in the Phoenix area, especially on many unprotected left turns such as highways. Sometimes, the vehicle cannot correctly understand basic road features, such as failing to adjust red and green lights to allow the car to more comfortably integrate into the speed of the highway." A woman claimed that she almost hit a Waymo vehicle because it stopped suddenly when trying to turn right. And it also encountered problems when encountering T-intersections.

Since the vehicles are programmed to strictly follow all driving rules, self-driving cars find it difficult to interact with human drivers who do not follow the driving rules. The following figure lists some of the complaints people have about Waymo's self-driving cars.

1. Acting too cautiously

At the T-junction closest to Waymo's Phoenix headquarters, Waymo driverless cars pass by every few minutes, turning left or right, sometimes several at a time. When turning left at this T-junction without traffic lights, Waymo driverless cars often run into trouble and cannot find an opportunity to merge into the normal traffic flow. The speed limit on this road is about 70 kilometers per hour. A left turn that is easy for human drivers to complete requires a long wait for the Waymo driverless car, which makes the human drivers behind it very upset.

Not only left turns, but also in places where you can't turn right directly or where there is no left turn lane, driverless cars are prone to making mistakes. Also at this T-junction, a lady drove a Jeep and prepared to turn left at this intersection, while the Waymo driverless car driving in the opposite direction was turning right. Unexpectedly, the Waymo driverless car on the opposite side suddenly stopped at the intersection halfway through the turn. Moreover, the driverless car did not turn on the turn signal to make it clear whether it was going straight or turning right. The two cars almost collided, and in the end the female owner had no choice but to bypass the Waymo driverless car and continue forward. In the picture above, the black X is where the Waymo driverless car stopped, and the black arrow indicates the route of the Jeep.

2. The problem of identifying a group of people

These vehicles have a hard time distinguishing between people and bicycles, and when they encounter groups of people, such as near shopping malls or parking lots, they will treat the data as a series of unrecognizable objects and have difficulty predicting where each person is going.

3. Parallel line problem

These vehicles often fail to reach the turning lane in time and miss the turn, which means they could have merged onto the highway but did not. An insider revealed that "this is because the gap search algorithm does not meet the requirements."

4. Two-way road without lane markings

East Jinxiu Street in Phoenix has two lanes in both directions, but there are no clear lane markings. Every time a white Waymo driverless car appears, it always drives in the middle of the road with a black sensor on top, which makes other vehicles on the road very uncomfortable.

In fact, some problems have long been recognized by the Waymo team. In an article published last September, it was mentioned that at the corner of South Landau Street and Southwest Avenue in Mesa, Arizona, there is a flashing yellow arrow that allows cars to turn left. Driving this type of intersection is tricky for both humans and self-driving cars. Drivers must carefully enter the five-lane intersection and then find a gap in oncoming vehicles. Turning left too early may cause driving hazards to oncoming vehicles; acting too late may mean making the driver behind the car very disgusted.

They propose that simulation can be used to turn a single real-world encounter (such as a flashing yellow left turn) into thousands of opportunities to practice and master a skill.

By using a powerful custom sensor suite, this complex left-turn virtual scene was built in Mesa, and by modeling the same size of channels, curbs, and traffic lights. Continuous reinforcement learning and repeated driving make the driverless car smarter. At the same time, by changing the speed, trajectory, and position of the object, different scene variants are constructed to train the driverless car, and finally pass the intersection smoothly in the real scene.

A Waymo spokesperson said the company is "continuously learning" and that "safety remains the highest priority" during testing. The spokesperson added that Waymo is using feedback from its early rider program to improve its technology, but it declined to comment on The InformaTIon's report.

The author believes that although Waymo has developed rapidly, it still has shortcomings in some marginal scenarios and is still not perfect in interacting with other road users. At present, most simulation tests are based on the driving ability of the unmanned vehicle itself. Is it necessary to introduce other human operators in the virtual world to follow the unmanned vehicle and feel the real state of unmanned vehicle driving from a third-person perspective? In addition, although Waymo's unmanned vehicle is close to mass production level, it is not good at the interactive language between the vehicle and the outside world. Ford's recent safety report mentioned that Ford's unmanned vehicle installed a row of indicator lights in the front row to help humans understand the status of the unmanned vehicle, and drive.ai also designed a device similar to an "LED prompter" on the body to indicate the status of the unmanned vehicle. However, Waymo does not seem to have considered this aspect so thoroughly.