Didi's Reinvention

Du Chen

When Li Jian thinks back on the code he spent long, sleepless nights developing for Udacity’s self-driving car competition, he can’t repress a look of elation. On the day of the competition, he and his team met with a group of engineers, who had been working in this field for years already. And what they saw in Li and his team’s work wasn’t just a bunch of code, but proof of their potential to join their ranks.

Those engineers were from Didi.

This year, Didi and Udacity hosted the first ever open source coded self-driving car race. Some 4300 contestants from the US, China, Russia, and the rest of the world took part, including no less than representatives from Stanford, Facebook, Google, Microsoft, and Tesla.

The “race” set a high bar: contestants had to create software packages using open source and self-produced code that could be ported directly to Udacity’s driverless test vehicle, and then operate it normally. But this was not a race you could win just with speed. The contestants had to incorporate automatic safety and sensor processing into their algorithms so that the car would be able to recognize pedestrians, other vehicles, and other obstacles, and then make in situ judgements that would protect the passengers, the car itself, and everything else on the road. After months of trials, a pair of Chinese programmers, Ying Zhenzhe and Li Jian, emerged in first place.

But from among all the contestants, Didi plucked out a few “diamonds in the rough,” young engineers who had clear potential to develop self-driving technology further. More importantly for Didi, it was able to gather data on the results from the competition that it could apply to its own efforts.

Recently, we paid a visit to Didi's Silicon Valley lab, where it has built up a team of more than 50 people and is still growing, to talk about precisely these kinds of experiments, and how the company is working to create new technologies that will reshape modern transportation. 

Self-driving technology is certainly a part of that, but not the whole. Didi's lab has been working towards a number of innovations, many less attention-grabbing, but more immediately workable, than driverless cars.  

At the company’s 2016 annual banquet, founder and CEO Cheng Wei called the Didi “the most impetuous company” in the world. It had developed too fast, but success was seemingly at its fingertips. Still, he was anxious. He believed that they still had not brought their user service and efficiency to where they should be. To put it plainly, their success was born out of subsidization. From that point on, Didi turned from its earlier fixation on swallowing up market share and began to commit itself to technological development. Now, Cheng is finally able to say that “Didi is a technology driven company.”

Those who hail a ride in Beijing today might find that they get a faster response than before, or that if a driver fails to show up for some unexplained reason and the user must cancel the ride, the system will automatically know not to fault the user. These are the subtle results of laborious optimization research and development on the technical side. 

The algorithm to discern who was really responsible for a ride being cancelled, for instance, began testing last December. From Didi’s perspective, it’s an important development, because it affects all of their users, both riders and drivers. Whether the cancellation penalty goes to the rider or the driver is determined by factors like expected drive time, the car’s realtime position, traffic conditions, and past cancellation patterns. So far, the machine learning model has reached 90% accuracy.

Didi says it is also extending its efforts to benefit those beyond its userbase. Starting in May, it worked with traffic cops in the city of Wuhan in installing a smart traffic system. Using route data from its own database, and ground sensor data supplied by the police department, it was able to model and reduce traffic under varying conditions.

The setup was simpler than it might sound. At a series of trial locations, 13 smart traffic lights were installed, their timing and sequencing varying according to the system. Didi says that it has reduced delays at morning rush hour by 30%.

Didi is not unique in this, of course. Uber tried something similar previously, using the data it had collected from rides in a selection of cities to create a platform for improving traffic management in those same cities. Didi, however, had by that point already deployed its system to twenty cities across China.

Last month, Didi brought Henry Liu, professor at University of Michigan, onto its staff as head scientist. Liu invented the SMART-Signal transportation system in use in California. He says that the caliber of research in industry is no less than that in academia, and with a company like Didi, where there are 20 million orders per day, there is a wealth of data and the excitement of huge potential for more directly transforming data into solutions.

Not long ago, Didi and a number of AI specialists at Stanford established a partnership, with Didi sponsoring several studies. Didi provides its research partners some 70TB of data per day from its ride orders, making it one of the most attractive partners for academics looking for big data sources.

Likewise, Didi shares its own research results.

For instance, Didi’s feature for guessing where you’re going is the product of an AI-powered predictive model, able to assess where a user is likely to want to go even before they type in their destination. It takes a mere two milliseconds to work, and it does so on an innovative algorithm developed by Didi engineers based on a normal distribution and Bayesian modeling. This destination prediction feature can reach an accuracy of more than 90%.

Cheng Wei has always believed that innovation in transportation has been essentially static since the first arrival of cars a century ago. According to Didi’s own estimates, in that time traffic speeds have remained consistently at an average of 30 kph, regardless of developments in engine technology.

Everything Didi has been doing has in some way been about answering that problem, trying to help push modern transit past the limits of internal combustion engines, parking space supplies, road congestion, and all the general inefficiencies in transportation. But they are not there yet. Which is why for Didi today, technology is all the more important, and not merely because market behavior requires it.

Its global strategy also has a role to play in this. Over the last two years, Didi has forged partnerships with companies that might otherwise be its competitors in regions all around the world, from 99 in Brazil to Taxify in Europe to Careem in the Middle East, not just for ordinary business, but to share, extend, and test its latest technological innovations. In turn, that has made the company’s Silicon Valley lab all the more important. Because Didi’s future will not be decided now by its ride-hailing services, but by what it can create that will move urban transportation beyond them.