What We Learned Driving an Autonomous Vehicle for 24 Hours Straight

By | January 9, 2018

Drive.ai’s First Drive-a-thon

At Drive.ai, we’re a team of over 100 people working to build the world’s best autonomous driving platform — to power a transformation in the way people, packages, and pizza get from one place to another.

Indeed, the world of mobility is ready for a change. In major metropolitan areas, a third or more of cars on the road on any given block are simply looking for a parking space [1]. In the world of tomorrow, however, individual vehicle ownership will plummet. Fleets of intelligent, autonomous vehicles will roam the roads, dynamically adjusting to the ebb and flow of the commute day. While our world is becoming more automated at a startling pace, the role of humans in this new transportation ecosystem will not disappear, certainly not overnight. So if we do want to build a fleet the can be on the road 24/7, what does that take?

Last month, we ran a 24-hour endurance event with the goal to see what it takes to achieve maximum autonomous vehicle uptime on a small scale. We had already driven at all times of day and in all kinds of conditions, but we wanted to further explore what it takes to run an autonomous fleet at maximum capacity. We called it the Drive-a-thon, and here’s how it went and what we learned.

The Drive-a-thon!

To start, we set ambitious goals: 400 autonomous miles, 90% autonomous uptime in a variety of different operating conditions, and most importantly, zero safety incidents. (We defined uptime as safety driver aboard, autonomous system engaged, doors closed, and vehicle on the road.) Our geofence of operation on urban and suburban streets had speed limits ranging from 25 to 40 mph, so hitting 400 miles would not be as simple as hopping on Highway 101-S and setting the cruise control. The whole event would take place on surface streets in Mountain View, CA: from 9am on November 16th until 9am on November 17th, 2017. We made it a company bonding activity; many employees spent the night at the office to cheer on the supporting cast.

We had one “primary” autonomous vehicle running continuously for 24 hours, and others coming on- and off-line as well throughout the event. Autonomous ride-hailing is a crucial component of our business model and a key piece of our technology stack, so successfully handling the pick-up and drop-off of some Drive.ai employees, friends, and family was another goal of the Drive-a-thon. We recently launched an upgraded version of our ride hailing app, and the Drive-a-thon provided a great opportunity for internally testing new use cases.

Sneak peak of Drive.ai’s ride hailing app

Today, it takes a lot of careful planning to keep even a single autonomous vehicle functioning for 24 hours straight. For starters, we are required by law to have a trained safety driver behind the wheel whenever we are on public roads in California. Our safety drivers demonstrate an even higher level of attentiveness and control than the average person on the road that is actually driving. However, one cannot expect a driver, supervising or otherwise, to be able to stay alert forever. For us, this meant shifts; we split the 24 hour period among five of our trained autonomous vehicle operators.

In addition to drivers, members of our mechanical team were on call back at Drive.ai headquarters, ready at a moment’s notice to attend to anything from a flat tire to a faulty sensor. Our infrastructure team was also on standby, having just installed an upgrade to our data logging system. One underappreciated aspect of autonomous vehicle operation is the data pipeline; our vehicles log multiple gigabytes per minute of sensor and telemetry data, for which we have ample onboard storage capacity for typical testing. For an endurance event, however, this meant “hot swapping,” or changing out logging drives without powering down the system.

Finally, we wanted this event to be fun. We invited the whole company, plus many friends and family, to come by: to keep the operations team company, to track our progress, and of course, to go on rides in our autonomous vehicles!

On the morning of our event, Murphy’s law of autonomous vehicle testing held true, and we awoke to a steady downpour of rain, which was forecast to persist throughout the day. Although we had plenty of experience testing in rain in the prodigious wetness from the start of the year, this would be our most thorough soaking to date. Nevertheless, we got right to it. At 9am, we were up and running.

Cruising along for a rainy commute. New safety driver, same rain

Interspersed throughout the day, we ran our own internal ridesharing app and completed dozens of rides, all while keeping our sights on the 90% autonomous uptime target.

As the evening rolled in, the downpour had lessened, and our driving world became a little drier and a lot darker. We passed the uptime halfway mark at 9pm, and were on decent pace to meet our goals, but knew that it was going to be a challenge to keep up the momentum for another twelve hours. Driving at night is more challenging from a perception standpoint. However, driving later into the evening meant that we encountered fewer cars on the roads — giving us a modest uptick in the distance we were able to cover per hour.

Tents around our office. Autonomous driving was even morein tents” than usual

When midnight rolled around, we decided to have some fun back at the office. We set up an indoor campsite for the all of our dedicated personnel that were staying overnight, ordered pizza, played games, and cheered on the operations team as they worked late into the night.

Snack break!

The wee hours of the morning were comparatively quiet on the roads (given the rainy maelstrom of the previous day), but held their own challenge. Keeping safety drivers vigilant and alert 20 hours in is crucial, so for each shift, we had at least one guest in the car, taking pictures, playing music, and enjoying our new autonomous in car UI and UX (blog post on that to come!).

Cruising late at night

As the sun rose, traffic picked up again, and we embarked on the third rush hour of our Drive-a-thon. During the final shift, from 7am until the checkered flag waved at 9am, it was amazing to see an autonomous vehicle run as smoothly in the twenty-fourth hour of operation as the first. For any vehicle, operating for 24 hours — with only one 5-minute stop for gas — is a lot. Despite the challenge, we were able to do it: keeping the sensors, computers, displays, and auxiliary systems alive and well for the entirety of the 24 hour event. A few minutes after 9am, we pulled back into the garage at HQ. The Drive-a-thon was a success!

The Results: By the Numbers

We had set aggressive goals for the event: 90% autonomous uptime, 400 autonomous miles, and zero safety incidents.

For the 24 hour period, our primary vehicle was in autonomous mode for 22 hours and 40 minutes, effectively 94% uptime. The main contributors to our downtime were changing out safety drivers (~35 minutes), swapping out data drives and related software maintenance (~25 minutes), addressing a software bug (~15 minutes), and one stop for gas (~5 minutes). When we passed the finish line at 9am on November 17th, we had driven 410 miles, meaning our average speed, including stops, was 17 mph for the event. Finally, and most importantly, we concluded our event safely.

Insights: What does it take to run an autonomous vehicle fleet?

The introduction of autonomous vehicles is transformative not only in how we move people and goods, but also for our world as a whole: land use, infrastructure, the insurance industry, wireless communications, and more will change radically over the coming decades. It’s not just about building autonomous vehicles, but also about developing the ecosystem in which they will operate. There are numerous considerations in running an efficient autonomous fleet:

  • Parking lots turn into Maintenance Centers: The average vehicle is parked for 22+ hours a day [2], while autonomous fleets will target 80%+ utilization. While this means a lower percentage of vehicles parked at any given time, the parking needs of an AV are different. Autonomous fleets will need specialized maintenance centers to serve as charging or fueling stops, data depots, sensor calibration sites, cleaning facilities, and more. We are in store for a dramatic repurposing of today’s paved paradise.
  • Pick-up and Drop-off: Autonomous mobility is hailed as a panacea for today’s “last mile problem,” but there are many aspects of a contemporary taxi experience that we take for granted. How do you make sure that the correct person has gotten into the vehicle? What if the person or people entering need help with luggage? What if an individual has a disability? This “last meter problem” is a critical consideration for fleets operating without humans at the helm. Negotiating a safe place to pull over for pick up or drop off is an often overlooked problem for an autonomous vehicle. Is it acceptable to block a driveway or a bike lane for a few minutes? If there is space, we need to detect the edge of the curb correctly, and ideally not use one that is painted red. If there is no legal space to stop, what do we do? Negotiating a contingency plan with a human driver is a common practice, but robots are still very new to this.
  • AV + EV: Electrification will no doubt play a huge role in the future of autonomous mobility. Energy cost for the hybrid platform we used in our drive-a-thon was about $0.09/mile, while a contemporary electric vehicle like the Tesla Model S is closer to $0.04/mile [3]. However, until charging and battery swapping technology matures further, hybrid vehicles will be at the sweet spot for maximizing uptime. Charging overnight is not an option for a vehicle at 90% utilization. Electric vehicle adoption has not yet fully taken off, but we believe that the introduction of AVs will be a major force in improving the infrastructure needed to support next generation energy storage technology. Businesses will be the ones who will really want this benefit, which is another reason Drive.ai has positioned to sell to business fleets.
  • Personnel: Today, many jurisdictions require a trained safety driver, either behind the wheel or supervising the behavior of the autonomous vehicle remotely. New jobs will be created for other tasks as well, such as: vehicle cleaning, general hardware maintenance, replacement and re-calibration of sensors from wear and tear, customer support, and more. Even in a safer, more automated world, we must still expect the unexpected, and many customers will prefer a human to assist in when issues arise. Will robots be able to empathize and react to unexpected situations as well as humans? This is an important consideration for us as we develop a user experience that builds trust and understanding with our customers.
  • Data: Autonomous vehicles generate gigabytes of data every minute, all of which must be accessible later for us to measure ride quality, track any critical incidents, and improve the technology over time. Some data can be streamed off-board via cellular, but most of it must be stored on-board, which means hard drives must be swapped regularly. The story hardly ends once the data is off the vehicle; it goes straight into our multi-stage data pipelines, which process data for future replays, debugging, visualization, annotation, and other analysis.
  • Human Robot Interaction: A passenger in an autonomous vehicle has opted into an experience where a robot is handling all aspects of the driving task. Inside the car, our riders can visualize information about what the car sees and how it is making decisions. Outside of the vehicle, however, interaction between our vehicle and its environment is a different story. Once we have removed the human driver from the picture, we no longer have the ability to make eye contact and emote with the outside world, at least not in the same way. We know that the transition from driven to driverless will not happen overnight. We have designed our systems to build trust in the future of autonomous mobility, not only with our customers, but in the heterogeneous world into which our technology must embark. Human-robot-interaction is yet another crucial consideration.

Looking back, our Drive-a-thon was a fantastic learning experience. Deploying an endurance event on a single autonomous vehicle gave us an excellent microcosm for understanding what it will take to keep a fleet of autonomous vehicles running at maximum capacity. And in 2018, we have our sights set on doing just that!

— — — —

  1. Shoup, D., and Campbell, “H. Gone parkin’.” The New York Times. <http://www.nytimes.com/2007/03/29/opinion/29shoup.html&gt; 2007.
  2. Morris, D. “Today’s Cars are Parked 95% of the Time.” Fortune. <http://www.reinventingparking.org/2013/02/cars-are-parked-95-of-time-lets-check.html&gt; 2016.
  3. Noland, David. “Life with Tesla Model S: One Year and 15,000 Miles Later.” Green Car Reports. <https://www.greencarreports.com/news/1090685_life-with-tesla-model-s-one-year-and-15000-miles-later&gt; 2014.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.