ARI SHAPIRO, HOST:
Self-driving cars are being tested in cities like Washington and London, places known for orderly traffic. But in Lima, a Peruvian startup is taking on one of the world's most chaotic driving environments, as Simeon Tegel reports.
(SOUNDBITE OF CAR HORNS HONKING)
SIMEON TEGEL, BYLINE: Peruvians have a saying. If you can drive in Lima - the Andean nation's capital - then you can drive anywhere.
MARIANA ALEGRE: (Speaking Spanish).
TEGEL: "Jumping red lights, refusing to stop for pedestrians and vehicles cutting each other up are all routine here," says Mariana Alegre, an urban planning campaigner. One study by price comparison website Compare the Market found Peruvians to be among the worst drivers in the world, second only to Thailand.
(SOUNDBITE OF CARS DRIVING)
TEGEL: Surprisingly, perhaps, that makes Lima an ideal location to trial the latest generation of autonomous vehicle technology.
ARTURO DEZA: It's extremely chaotic. There's probably about a hundred people around us. There's a policeman telling us to break the rules, too, and there's a lot of informal commerce, people selling fruit, clothes on the street and very tight spaces.
TEGEL: Arturo Deza is the founder of Artificio, a Peruvian driverless tech startup. A former researcher at MIT and Harvard, he is testing how this visual language model driving system, or VLM, reacts when one of his team drives through an informal market, blocking most of the street in downtown Lima. Essentially, the unpredictability of Lima's traffic can accelerate the self-driving AI's learning curve, as Deza explains back in the relative calm of the lab.
DEZA: There's a jargon word for this in the AI community that's called adversarial training, which means training a system in a more robust way through data points that are very unusual, and these would be cases in which people are breaking the rules.
(SOUNDBITE OF CAR HORN HONKING)
TEGEL: So we were just in standstill traffic, and a woman with her two young kids was jaywalking, crossing a street just behind us and almost got hit by a motorbike that was coming up between the cars.
Bhuvan Atluri is a professor of mobility at MIT. He says that minor road accidents and injuries happen between 20 and 200 times more frequently in developing nations like Peru than the United States.
BHUVAN ATLURI: Using data from other geographies with different edge cases, different traffic mode interactions, different norms and behaviors can greatly increase the pace and the learnings of training.
TEGEL: That explains how Deza, a neuroscientist who specializes in computer vision, is in the process of raising $7 million from investors for Artificio. The plan is to eventually let an autonomous test car navigate Lima's mean streets without a human driver. But first, the system needs thousands more hours of training, as Deza explains while reviewing footage from a dashcam test.
DEZA: It can identify the people. What it can't identify is the semantic, or the situation of, hey, I should be careful because this is a hazardous situation. There's a lot of cars. There's a lot of pedestrians, and there's children.
(SOUNDBITE OF CARS DRIVING)
TEGEL: But not all of Lima's lessons will be universal. The ultimate goal is for a single AI system to drive anywhere in the world, adapting to local conditions. To do that, it will have to read other drivers' behavior and understand that in, say, Chicago or Atlanta, it must drive like an American, not a Peruvian.
For NPR News, this is Simeon Tegel in Lima, Peru. Transcript provided by NPR, Copyright NPR.
NPR transcripts are created on a rush deadline by an NPR contractor. This text may not be in its final form and may be updated or revised in the future. Accuracy and availability may vary. The authoritative record of NPR’s programming is the audio record.