DISRUPTIONHUB interviews Ben Peters from FiveAI
If autonomous systems are going to work, we need to mend mapping. You’re on your way to meet an old friend for dinner, but you’ve never been to the restaurant. Your first instinct is probably to consult Google Maps. In 2017, the navigation tool was the seventh most downloaded app, and the most popular mapping app overall. Due to the expansion of self drive technology, it’s not just humans that need to know how to get from A to B. The maps that these systems require must be far more complex, using in depth geospatial data to predict constantly changing routes and environments. Which businesses are on the road to fixing navigation networks, and what does this mean for the driverless future?
The nuances of navigation
Google may lead the pack when it comes to finding a venue in an unfamiliar city, but not necessarily when it concerns the vast detail needed to enable autonomous technology. These systems need to compute an exhaustive list of changeable circumstances, including traffic, weather, road conditions, other vehicles and pedestrians. Humans do this using their senses, but building artificial awareness is far from innate. So how do you begin to direct driverless technology? It starts with a prior map. Automakers like Ford and GM are building their own maps for use in driverless cars, as are transport apps like Uber and Citymapper. Tesla is quietly working away at improving Autopilot, its advanced driver assistance feature. One of their partners is Mapbox, founded in 2010 to offer an open source mapping platform.
Another approach is building autonomous services themselves. Rising software company FiveAI wants to solve Autonomous Vehicle (AV) problems in the urban environment by doing just that. The startup has received a £12.6m grant from Innovate UK to trial their prototype autonomous service in London late next year, and expects to launch a public service in 2021. According to Ben Peters, from FiveAI, the real problems lie in navigating downtown urban environments.
“In a highway environment, everyone’s going in the same direction, and you can fairly easily model the behaviour of the actors. But when you move to an urban environment, pretty much anything could happen,” he says. “You’re never dealing with certainties, but rather a distribution of possible futures. Therefore, you frame it as an optimisation problem, searching through those potential futures for the path that gets you where you want to go with the least possible risk of causing danger.”
For the last decade, companies have relied on prior maps to tell them where they can drive. However, no prior map should be expected to be 100 per cent accurate. But how do you build a map that is suitable for driverless technology?
“Globally, there is no company that has yet solved complex urban environments from a level four perspective. To solve it, and safely, you need to apply a vast number of sensors, many cameras, many lidar, many radar units around the vehicle, and you need a huge amount of computing,” explains Peters.
It’s clear that the major challenges to building navigation networks include the huge amount of time and resources needed to deliver constant updates. This has led many auto incumbents to work with traditional mapping companies. Another issue is unpredictability. No matter how much predictive technology a company has, there will always be anomalies. Spotting them, and responding safely, is perhaps the biggest obstacle for transportation firms. Aside from these general considerations, other potential setbacks depend on the product that a company wants to deliver.
“For car companies building vehicles for consumers to own, that’s a car with a lower level of automation like Level 2 or Level 3, generally speaking the map is considered to be ground truth and the focus is how they make sure maps are always up to date and that they can localise vehicles on them in any conditions,” says Peters, “If it’s a geospatially limited service fully autonomous, Level 4 solution like us, there are multiple potential sources of truth. If your map says one thing and your computer vision says another, which is true? There are ways in which you can figure out which system to trust. For real time perception systems, it can be broader contextual knowledge. For example, if you know there’s snow, your real time data detection will be worse than on a light, clear day.”
At this relatively early developmental stage, AV technology relies largely on maps to function safely. As a result, the creation of advanced real time maps is likely to increase the reliance on location services and open data. For example, autonomous vehicles are going to need to know exactly where pedestrians are, all of the time – as is a delivery bot carrying a hefty piece of machinery across a street. Provided that a company can trust a map – which means applying context and cross referencing with visual perception software – they will be imperative to facilitating the widespread use of autonomous vehicles. For FiveAI, building reliable maps that work in conjunction with artificially intelligent computer vision could be the key to solving congestion in European cities.
“If you’re trying to deliver a service within an urban environment, you’re likely to be drawn down the route of creating your own maps and just using open data as a prior,” says Peters, “But traditional mapping companies like TomTom have a large vested interest in gathering, mapping, and maintaining map data, and then delivering APIs on top for customers. This is probably the right way if you’re using the maps for ADAS features, for lane information, for example where you care about having the largest number of roads mapped and maintained to a high frequency and accuracy.”
‘X’ marks the parking lot
One day, says Peters, autonomous vehicles will no longer rely on prior maps. Instead, they will use the way-point guidance that comes naturally to human navigators.
“The reason we rely heavily on prior maps at the moment is because the most important thing is safety. Real time perception systems have multiple failure modes, and some of those failure modes can be mitigated with prior maps,” he explains. “Prior maps are important, but that won’t be the case forever. At some point, the idea is that these machines will be able to perform as a human would do and have a waypoint based system.”
Maps have already undergone fundamental disruption. Ask a millennial to get hold of a map, and they’d be more likely to visit a webpage than a newsagent. Travel needs are changing to accommodate the growth of driverless technology, whether that be an industrial robot or an unmanned drone. Regardless of whether a company is working on an AV model or a shared autonomous service (like FiveAI), the challenge is to draw together vast quantities of real-time data to help autonomous systems navigate any environment safely.
“We’re still a few years away from having all the answers,” says Peters, “But I can’t understate how transformative autonomous technology will be.”
Can Google retain their lead in maps? Will automakers and software companies compete or collaborate? How might inner city transport be affected by autonomous, shared transport services? Share your thoughts.