On the road in San Francisco and Phoenix with 7 new graduate engineers

Raring to go in San Francisco.

Fully autonomous taxis may not be a reality in Japan yet, but in cities like San Francisco and Phoenix, they’re already on the road. During our trip, we hailed robotaxis via apps for rides to the airport, our hotel and the supermarket. We also took a Tesla vehicle for a spin on city streets and highways to test its autonomous driving mode. Seeing a driverless taxi pull up at the tap of a button is an experience that most people in Japan are yet to enjoy.

 

A Waymo vehicle arrives for a pickup in Phoenix.

During the trip, we rode in robotaxis by Waymo and Cruise, and tried a beta version of Tesla’s Level 2 Full Self-Driving feature in a rental vehicle. We analyzed the experience from a technical perspective, focusing on perception, path planning and user experience.

 

During tests in a Tesla, the driver tentatively lets go of the wheel with the vehicle running in Full Self-Driving mode.

Perception

Sensor overview

First, we’ll take a look at autonomous driving systems, focusing on sensors, environmental perception and object recognition. The Waymo vehicles we rode in varied depending on the area. Jaguar I-Pace vehicles served the area around the airports in San Francisco and Phoenix. The vehicles were equipped with the 5th-generation Waymo Driver system.

The sensor configuration included a top-mounted LiDAR on the roof for 360° sensing of the surroundings, with cameras positioned around the vehicle. A long-range camera and radar system handled forward long-distance sensing. Four LiDAR, camera, and radar units were mounted at the front, rear, and just above the front wheels to detect objects close to the vehicle. A camera and radar system was also installed at the rear, just below the turn signals.

In another part of Phoenix, we rode in a Chrysler Pacifica Hybrid. The vehicle was equipped with a top-mounted LiDAR, camera, and radar system, along with LiDAR and radar units just above the front wheels, additional LiDAR sensors at the front and rear, and a radar sensor mounted at the top rear. It is an earlier model than the Jaguar I-Pace vehicles, and the UI of the in-vehicle system was significantly different.

 

Perception accuracy

One of the most interesting aspects of the Waymo ride was how rarely it made perception errors. The vehicle displayed perception results on a display in the vehicle, and while the detected objects sometimes appeared to slightly change in size, there were no cases of oversights or misidentifications. This was especially notable in San Francisco, where we rode in heavy rain and at night. Even in these challenging conditions, the system consistently detected pedestrians and other vehicles.

I had watched several videos of Waymo journeys on YouTube and official demo footage before experiencing a ride firsthand. Even though the system rarely failed to detect objects in those videos, I’d expected a few hiccups. However, I was impressed by the accuracy. Also impressive was the wide range of objects it could recognize. As well as pedestrians, motorbikes, bicycles and cones, it also identified things like kickscooters and trash cans. I was particularly surprised when the Waymo vehicle slowed down after a cat darted into the road from between parked cars at night.

 

Path planning

Waymo's control was very smooth. It seemed to always consider the trajectory several dozen meters ahead, avoiding unnecessary steering corrections and unnecessary acceleration or braking. Even when changing lanes, the rate of longitudinal and lateral acceleration felt natural, almost as if a human were driving. I was especially impressed by how smoothly it released the brakes, maintaining a nearly constant deceleration, and stopping at the intended position.

 

Planning in lane-free zones

Another interesting aspect was Waymo’s planning performance in areas without lane boundaries. In a supermarket parking lot, the system's route planning became unstable, with sudden changes or interruptions, which might indicate a bug. However, during maneuvers like turning around after reaching a dead-end or pulling over to the side of the road, the vehicle was able to execute a U-turn smoothly. I’m particularly curious about how planning in open spaces is integrated into the system.

 

Predicting oncoming vehicles at intersections

At an intersection without traffic lights, our vehicle was preparing to turn left when we encountered an oncoming car on the opposite side of the intersection that seemed to be going straight. Since vehicles going straight have the right of way over those turning left, our vehicle waited for the oncoming vehicle to pass. However, the oncoming car turned on its turn signal just before making a right turn. Our vehicle followed after the other vehicle had turned, which suggests the Waymo system accurately recognized the turn signal.

 

An oncoming car waits at an intersection.

MRM and autonomous recovery

The Cruise vehicle was able to continue driving without issue, even in heavy rain or hail. When abnormalities are detected while driving, such as from false detections, the system is designed to operate what is known as a minimum risk maneuver (MRM) and pull over in a safe spot. We experienced such a situation, and once everything had returned to normal, the vehicle resumed autonomous driving.

 

Our vehicle used a minimum risk maneuver to come to an emergency stop, as depicted in this illustration.

Unfamiliar territory

In the U.S., at a 4-way stop intersection without traffic lights, the rule is that the vehicle that arrives first has the right of way, and when crossing at right angles, drivers must yield to the vehicle on their right. While driving through downtown San Francisco, we encountered a situation where our Waymo vehicle was in a position that should have given it the right of way, yet it yielded to a truck coming from the left. This made me curious about how Waymo handles decisions regarding arrival order and priority when making turns.

 

Our vehicle gives way to a van approaching from the left.

With Cruise, the vehicle would come to a stop at speed bumps, which could lead to occasional honking from following cars, especially during the day. In contrast, Waymo would simply slow down at speed bumps. I also noticed that Cruise sometimes had difficulty with pullover maneuvers. When requesting a pickup via the app, it would often approach the location but then move a few meters further before returning. I found myself chasing after Cruise a few times in the rain, which was an interesting experience.

Autoware also has a parking feature, and we are continually working on improvements to better support regions like the U.S., including MRM integration for enhanced safety and functionality.

 

UI/UX

In-car and external interface

I was surprised that neither Waymo nor Cruise were taking advantage of displays or voice systems for interactions with people outside the vehicle. Inside the vehicle, Cruise only showed the current location and route to the destination, while Waymo displayed additional information, like why the vehicle had stopped, which helped provide reassurance. It also seemed like Waymo invests a lot in monitoring passengers inside the vehicle, reflecting a strong focus on safety and operational management.

Both Waymo and Cruise seem to have systems in place to prevent major accidents, using multiple cameras to monitor vehicles and operators who can communicate with passengers if anything goes wrong.

 

Screenshots of the taxi-hailing apps we used (right: Waymo; left: Cruise).

Passenger experience

Riding in Waymo and Cruise vehicles was a smooth experience. The driving was so precise that any initial concerns about potential collisions quickly faded. Waymo displays perception results on vehicle displays, which helps to put passengers at ease. The limited operating range and occasional difficulty with stopping were minor drawbacks, but aside from that, they functioned like regular taxis – so much so that by the 10th ride, the novelty had started to wear off.

Tesla’s FSD performed well as a driver assistance system, and the ability to use it anywhere was a clear advantage. If it made a risky move, switching to manual driving was always an option, which provided some peace of mind. It also includes features like following traffic cones and automatic parking, but we didn’t get the chance to try them on this trip.

Other observations

Pickup point

Cruise vehicles often stopped tens of meters away from the designated pickup point, which is something that would rarely happen with a human driver. This issue may stem from the prevalence of street parking in the U.S. Autonomous vehicles assess whether they can safely stop at a designated location based on the surrounding environment. In areas with heavy street parking, it can be difficult to find a spot that meets the criteria, causing the vehicle to overshoot the pickup point.

In contrast, Waymo appeared to handle such situations with much more consistency, even in areas with heavy street parking, and incidents like this were rare. At TIER IV, we are actively developing street parking functionality and continuously working to improve performance.

 

A Cruise vehicle pulls over on a residential street with lots of parked cars.

Highway hurdles

In the autonomous driving industry, it's widely acknowledged that urban environments pose greater challenges than highways. However, based on my experience as a passenger, I couldn’t help but feel that when it comes to deploying unmanned services commercially, highways might actually be the tougher road to navigate.

Urban driving seemed well-developed across services, especially with Waymo, which performed at a level comparable to human drivers. The experience was smooth and reliable, with minimal disruptions. Even when issues did occur, they were minor. In some cases, the vehicle came to a stop, causing other drivers to honk their horns.

However, despite this level of refinement, Waymo has yet to expand its service to highway speeds. Tesla offers a hands-free driving function on highways, but situations such as merging onto highways can feel scary due to the higher speeds. While robotaxi services are set to expand, the progress of robotruck operations is less obvious, which may reflect the additional challenges linked to highway deployment.

One possible explanation is that the biggest challenge in deploying unmanned services isn’t just navigating complex environments – it’s the difficulty of designing fail-safe systems. Fail-safe design ensures that, in the event of an emergency, the system can respond safely.

A useful comparison is the aviation industry. Airplanes face fewer operational scenarios than autonomous vehicles, and automation technology has been widely researched. Yet, fully unmanned passenger flights remain rare, likely due to the immense challenge of designing fail-safe mechanisms that meet the necessary safety standards.

For highway driving, this suggests that purely technical improvements to reduce emergency scenarios may not be enough. Compared to urban environments, regulatory frameworks and industry standards may play an even greater role in enabling safe deployment.

Wrap-up

This post has been a summary of our experience riding autonomous vehicles in the United States. In some ways, the companies developing the technology are both allies, driving efforts to expand the market, and rivals to surpass. TIER IV leads the development of Autoware, open-source software for autonomous driving. We believe open-source innovation will play a key role in shaping the future of mobility, and we’re always looking for people who share that vision. If you’re interested in joining the journey, get in touch.