In the tropical midday heat, I squint to take in the two-hectare test course before me. Here at Singapore’s autonomous vehicle (AV) test center, there’s no shortage of ways to put driverless vehicles through their paces—with slopes, crosswalks, bus lanes, traffic junctions, and even a crank course consisting of sharp 90-degree turns.
But it’s an unassuming structure at the far end of the center that catches my eye. There stands a thin metal frame nearly 40 meters long, split into nine sections. It looks like a series of empty door frames. Atop each section sits three nozzles, together capable of releasing up to 150 milliliters of water per hour.
It’s a rain simulator—one of the few found in AV test centers around the world. CETRAN, or Singapore’s Centre of Excellence for Testing and Research of AVs, also has a flood simulator. And for the first time since the center’s opening in late 2017, those simulators will soon be put to use as two companies, including MIT startup nuTonomy, begin adverse weather testing of their AVs in the coming months.
“We do not have [AV] sensors that can function properly in heavy rain and flood conditions yet,” says senior scientist Niels de Boer, who runs CETRAN. All AV companies have to test their vehicles at the center before being allowed onto public roads. So far, most have prioritized functionality and safety over adverse weather testing, he says.
The tropics, where sudden downpours of heavy rain frequently lead to flash floods, present special challenges for AVs. “Detecting standing water when it floods is difficult,” says de Boer. “And even if you can detect the water, you don’t know how deep it is.”
When it rains, water droplets can absorb laser beams and radio waves, produced by lidars and radars respectively, and cause signal attenuation. Or the droplets may reflect these waves and wrongly register them as an obstacle in the vehicle’s path, says engineer Gil Jr Opina, who works on developing autonomous electric buses and other vehicles at Singapore’s Nanyang Technological University (NTU).
Rain can also distort the images collected by a camera and “generate atmospheric veiling effects similar to mist or fog,” adds roboticist Daniela Rus from MIT who is working on the Singapore-MIT Alliance for Research and Technology’s fleet of self-driving cars and golf buggies. “All these degradations make it difficult to detect static and dynamic obstacles, resulting in poor localization and navigation.”
To overcome these challenges, developers often seek solutions in software. When creating reference maps to help vehicles navigate, Rus and her team select “permanent or long-term features for localization that are robust to the changes in weather conditions.” These include tall buildings, traffic poles, and tree trunks. “If it’s a tall building, even if the rain obstructs a few data points, you can still get information from other points on the building,” she says.
To correct for poor camera vision during heavy rain, Rus’s team applies de-raining and de-hazing filters, trained using deep learning techniques, to the live video stream. Meanwhile, Opina and the NTU team extract information collected by the vehicle’s various sensors, use machine learning to correct for any poor signals obtained, and then adjust the algorithms that control the vehicle’s behavior. In response, the AV may slow down or alter its data collection rate.
To further develop a comprehensive testing program for AVs in all types of terrible weather, Singapore’s CETRAN wants to collaborate with other AV test centers overseas such as The American Center for Mobility in Michigan, AstaZero AB in Sweden, and K-City in the Republic of Korea, says de Boer.
To say that Singapore’s government is pushing hard to develop the nation’s AV sector would be an understatement. Earlier this year, the city state released national standards to guide the development of driverless vehicles—making it possibly the first country in the world to do so. In February, Singapore beat 18 countries, including the United States, to clinch second place in the Autonomous Vehicles (AV) Readiness Index [PDF] published by consulting firm KPMG International.
“I think one aspect that makes Singapore unique is that it’s really a government-initiated activity,” says de Boer. Comparatively, large automobile companies and tech giants are driving the push for autonomous vehicles in other countries, he says. Think BMW in Germany, Volvo in Sweden, or Waymo in the United States.
Instead of private cars, Singapore envisions autonomous shuttles, buses, taxis, and even road cleaners. With 5.6 million people living in an area fewer than 725 square kilometers, or roughly two-thirds the size of New York City, Singapore is the third most densely populated country in the world.
Transportation infrastructure, including roads and parking, occupies close to 20 percent of Singapore’s total land area, says Anshuman Tripathi, who leads AV research at the Energy Research Institute at NTU (ERI@N). Plus there are now “more and more rush hours” as the number of journeys made by residents has increased by 15 to 18 percent in recent years.
Public transport AVs, if introduced, could support a “car-lite” society and ease congestion. They would also bring the country one step closer to its vision of becoming a so-called Smart Nation, a goal announced five years ago.
In the coming months, AV trials in Singapore will begin to move away from university campuses and other controlled environments, into residential areas. If all goes well, Singaporean commuters could see the first AVs for public transportation deployed in three small towns by 2022. Says de Boer, “We’re pushing very, very hard to achieve that goal.”