Data scientist Claire Lebarz joined forces with four other Turo employees to take on the Prius Challenge, where they hacked a Prius to maximize efficiency and competed against 20 other teams. Here’s their journey to hack the Prius.
THE SPARK THAT LAUNCHED 1,000 PRIUSES
It all started earlier this year at the CES in Las Vegas, where Toyota unveiled the CONCEPT-i. Gill Pratt, CEO of the Toyota Research Institute (TRI), shared his vision of a future where driving remains a lot of fun, and AI supports the driver to enable smarter driving. This notion piqued my interest.
A week or two later, Google Launchpad hosted a CES follow-up meetup, where Stephen Hughes of TRI shared his thoughts about the future of connected mobility. He invited the audience to participate in the first open edition of the Prius Challenge, where competitors attempt to maximize the efficiency of a Prius over nine laps in 42 minutes at the Sonoma Raceway. Now I was very interested.
The idea of mixing analytical and driving skills seemed like a really cool way to integrate the theoretical and the practical, and I was certain I could easily convince some of my colleagues to join me on this adventure. I also liked the fact that, for once, it wasn’t about being the fastest on track, but being the most efficient; it was about optimizing for something meaningful.
Research & practice
I had never really wondered how a hybrid car works — sure, I had heard that people tend to drive them slowly and that hyper-mileage websites existed, but I was definitely not an expert. As a seasoned data scientist, however, I know that most value from data comes from meaningful transformations and assumptions, and most of all, from asking the right questions. The Turo team — consisting of two engineers, two data scientists, and our head of airport operations — all needed to get into the car and start understanding how it works.
So we rented a Prius from Ayako and her husband Ezra and played around with it — recharging the battery, figuring out dos and don’ts, etc. We realized that by going downhill with a high enough speed, the regenerative braking was recharging the battery efficiently; braking at stop signs or red lights in the city was not moving the needle much. Not surprisingly, we realized smoothness with the throttle was very important, and we now had a sense of how much of a difference that could make: 10 to 20 mpg in the city. It became clear that the electric engine could not take us all the way on the big San Francisco hills, so we took the car to the highway, where we learned that building and keeping momentum was key to tackling uphill efficiently.
Scheming & strategy
To help us prepare for the challenge, TRI opened access to data from 27 Sonoma Raceway laps of professionals driving the Prius. In addition to the track layout — latitude, longitude, and altitude — we had access to the vehicle speed, engine speed, throttle, fuel consumption, driving mode, and battery state of charge. The most efficient lap achieved was close to 150 mpg — quite impressive! With this new understanding of the car, it was time to analyze the data, most of which made sense.
Our approach was to conceptualize the challenge into two optimization problems:
- The best lap strategy given a state-of-charge of the battery
- The optimal battery management across the nine laps
We identified that we could further break down a lap into three rough sections: the uphill challenge (from turn 1 to turn 3a), the downhill challenge (from turn 4a to turn 10) and flat U-turn challenge or turn 11a.
The uphill section (darker blue in Figure 1) appeared to be highly dependent on the battery charge and time, while the rest of the lap enabled us to recharge the battery for the next lap through regenerative braking.
So the dominant strategy was to build momentum in anticipation of the uphill. On the flat and mostly straight segment of the road, we would accelerate and then let up on the throttle once we had accumulated enough speed to coast in Neutral on the remaining part of the hill. (See blue vs. red strategy in Figure 2). With high enough charge (above 71% in our case) and small enough uphill (turn 3 to 3a), EV mode would be sufficient all the way to the top (see green strategy in Figure 2).
The two last sections were mostly tradeoffs between recharging the battery (Drive or Brake) vs. free miles (Neutral) at higher speed. Downhill, Drive was the best way to recharge the battery when we needed it, and we preferred Neutral for turns and straight lines, because it equated to free miles and removed engine friction.
In order to test out these strategies, we rented a few Priuses for practice. We couldn’t access the Sonoma Racetrack, so we decided to hit the highway. Each driver’s performance rendered very different results, and I began realizing that each driver presented with a different challenge to execute the strategy:
- Jérôme excelled at throttle smoothness, but was having a hard time accelerating.
- Anton was the opposite. He could change speed and driving style rapidly, at the expense of smooth transitions.
- Pushkar was used to race driving, which made him comfortable and familiar on the track but he could get frustrated by the slower speed at which the Prius should be optimally driven (35 mph on average).
- Doug was our electric car expert, and the most experienced driver. For Doug, safety (rightly) comes first, but makes him respond slower to instruction.
All of this made plain the need for a simple strategy for race day that we could adapt to each driver, as well as to traffic — since there would be nine other cars driving on the circuit under the same conditions.
On the day of the challenge, the team operated like a well-oiled machine, and we kept our strongest players in their respective roles for the bulk of the competition. Anton and Doug served as the pit crew, monitoring lap times and competitor performance. They relayed all of this info to Pushkar, who was in the back seat, while I served as Chief Strategist in the front passenger seat, advising Jérôme, our primary driver. Jérôme and I spoke in French (our native language), which comically left the event-assigned umpire completely out of the loop.
During the race, we did not adapt fast enough to live data and intuition. As early as lap 2, we realized we would not be able to reach 60 mph by turn 1 with smooth acceleration. Instead, we focused on improving execution with our smooth driver, accumulating a time penalty. We had to move away from our theoretical strategy and adapt to live data.
When it was all over, we finished just ahead of the middle of the 20 competitors for mpg — we did our first lap at 150 mpg (the second best in the challenge) — and given the high caliber of competition, this was an achievement.
Maybe the greatest lesson was that the best theoretical approach was not always the best in practice. While a successfully implemented uphill strategy can theoretically achieve a 300 mpg lap, the challenge for a person executing it — when you take into account complexity and traffic — is extremely high. The strategy was risky because poor execution — not gaining enough speed on the steepest part of the hill — could lead to under 80 mpg. Meanwhile, a simple strategy of maintaining constant speed of 35 mph on the first section could have guaranteed 85 mpg lap.
As it turns out, the theoretical optimal strategy isn’t necessarily achievable by a real live driver, but it can be adapted to his or her driving pattern. We’re very excited to develop a human/machine learning approach to this challenge, in which the computer’s strategy learns from human driving patterns and the driver improves their driving style based on computer recommendations. We’re already planning our tactics for next year, when we take on the Toyota self-driving Prius!