The experts at atlatec are specialised in the creation of maps for autonomous vehicles. Here they tell more about the importance of high-resolution maps for autonomous driving, localisation and simulations.
This July a Waymo press release made quite the splash in the autonomous driving scene: The Google spin-off has now driven over 10 million miles in the real world - and over 10 billion miles in simulation.
Why does the autonomous driving powerhouse simulate 1,000 x more miles than they actually drive? The answer is not simply "because we can" as you might expect, given Alphabet's computing power and expertise. Waymo has made it clear that simulation is key to their autonomous driving achievements in real life: CEO John Krafcik stated in 2017 that 80% of algorithm improvements come from simulation.
A look at Waymo's performance on their mission to "building the better driver" gives credence to this approach. Their disengagement rate is down to 0.09 per 1,000 miles driven: Only every 11,154 miles (17,951 kilometers) does a situation occur where a human driver needs to take over.
An impressive learning curve that others have not been able to follow: Cruise and Zoox reach half or a quarter of that distance, respectively. Other players barely register on the scale set by Waymo, many clocking in at single-digit distances.
There is disagreement whether autonomous vehicles should use high-definition maps at all: Some feel that production vehicle sensors can become powerful and reliable enough to process every environment in real time. One prominent representative of the sensors-only camp is Tesla’s Elon Musk who has called high-precision maps “a really bad idea.” Others disagree and believe that Tesla’s fatal accidents might have been avoided if their autonomous driving system had used a map as redundancy fallback for unclear driving situations.
"But no matter where you stand regarding high-definition 3D maps in production vehicles," says Dr. Henning Lategahn, CEO of mapping company atlatec, “Waymo’s success makes it clear that for simulation there is no substitute for real-life data. A digital twin of an actual roadway is closer to reality than any artificially created map.”
Optimising for real-life autonomous driving conditions is easier if you build simulations on real-life data. Not only does it save the effort of designing complex environments from scratch, it also surprises you with scenarios you might not think to include - like, say, a pair of geese crossing your path in the middle of town:
Simulations also have the advantage of total freedom: Manufacturers can test their vehicles and components in extreme situations and at scale - without having to worry about regulation or bad publicity as a result of possible accidents. This trend seems to be catching on: A recurring request for Henning’s team at atlatec is the mapping of proving grounds. “After we’ve captured and processed a route once, our customers can re-use the map without limitations. Additionally, they can re-create identical environments to compare different vehicle configurations.”
Whether it’s about exploring autonomous cars, remote-operated trucks or new driver-assistance systems: Conditions on the mapped routes never change (unless the client requests an update), so they allow for a broad range of precise A-B tests or long-running experiments, all within the simulators automotive companies already use.
At Automotive World’s 2019 M:bility Europe conference representatives from SBD Automotive, Blickfeld, HERE and Baselabs discussed the challenges different approaches face:
On one hand, sensors don’t always work perfectly, for example in extreme weather situations or foggy environments. On the other, mapping the world’s 20 million miles road network (32 million kilometers) seems an incredible challenge - and still requires sensors for localization and situations like weather-specific driving rules. Considering that different data sources might provide conflicting data/information, it seems crucial to know when data is unreliable - and to have a functioning “error model” in place.
The panelists amicably reached common ground, though - and Henning agrees: “For the foreseeable future, there is no either/or question for autonomous driving. The best solutions will combine the advantages of highly performant sensors and HD maps.”
atlatec builds high-definition 3D maps for use in autonomous driving, simulation and ADAS. The German company uses a very lean sensor setup, enabling them to collect road data all over the world on demand and in short time. The data processing and annotation work is done with an entirely self-developed software pipeline, using a combination of artificial and human intelligence.
The company’s maps are used by OEMs and suppliers around the world such as Ford, GM, Daimler, Volkswagen, Bosch, Continental, ZF and others.
atlatec will be part of New Mobility World 2019 and can be found in hall 5.0, booth B3408.