Autonomous Vehicles vs. Kangaroos: the Long Furry Tail of Unlikely Events

Self-driving cars in Australia are preparing to handle kangaroos, but what about autonomous cars everywhere else?

4 min read

Evan Ackerman is IEEE Spectrum’s robotics editor.

A small kanagroo stands illuminated in the headlights of a car
Photo: Theo Allofs/Minden Pictures/Getty Images

One of the biggest challenges for autonomous cars is being able to deal with potentially dangerous events that don't happen very often. This is why most autonomous car companies are spending an enormous amount of time and effort driving around and collecting sensor data: they’re hoping to expose their algorithms to as many unusual things as possible, to give them more experience with making decisions even when weird stuff happens. “Weird stuff that's hard to deal with” can mean almost anything, but it tends to mean different things in different places. In San Francisco, it probably means bicycles. In Sweden, it means moose. In Australia, it means kangaroos, and according to Volvo, kangaroos are proving to be very weird stuff indeed.

Most autonomous cars do have animal detection systems, but they’re primarily designed for detecting medium to large quadrupeds, like cats, dogs, and deer. Kangaroos not only don’t look like quadrupeds, they don’t act like quadrupeds either, and all of that charming hopping about really does a number on an autonomous system trying to figure out where they are and what they’re doing, as ABC News Australia reported last week:

It turns out the unusual way that kangaroos move completely throws off the car's animal detection system. “We've noticed with the kangaroo being in mid-flight ... when it’s in the air it actually looks like it’s further away, then it lands and it looks closer,” Volvo Australia’s technical manager David Pickett said. Because the cars use the ground as a reference point, they become confused by a hopping kangaroo, unable to determine how far away it is.

Furthermore, kangaroo identification is made more complex because of how different kangaroos look depending on what they're up to. To an autonomous system, a standing kangaroo looks different than a grazing kangaroo, which looks different still than a hopping kangaroo. What’s more, any given kangaroo can rapidly (and sometime erratically) switch between these three states. The upshot is that you need a system that's able to reliably identify kangaroos despite this variability, which is a very complicated thing to put together. Apparently, Volvo sent a research team to Tidbinbilla Nature Reserve in Canberra 18 months ago, specifically to study kangaroos. They’re still working on solving this “quite interesting” problem.

The kangaroo problem illustrates several potential issues with autonomous cars. First, there's localization—not the “where am I” localization, but rather the “being prepared for a new place” localization. Practically, this involves tweaking software to be able to recognize and handle the unique characteristics of a unique environment, and may involve differently oriented traffic lights, new types of street signs, foreign languages, completely new rules of the road, or (in the case of Australia) marsupials with cyclically variable altitudes.

Secondly, kangaroos are, everywhere but in Australia, a good example of autonomous cars' long tail problem (heh), which suggests that as the number of possible driving situations you may encounter approaches infinity, the probability of encountering those situations approaches zero. In other words, lots of weird stuff could potentially happen, and there's no way to predict all of it. If you're driving around in the United States, encountering a kangaroo would be very weird, but it's by no means impossible: here's one that was running around eastern Oklahoma in December of 2013:

As it turns out, that particular kangaroo was not Lucy Sparkles, a kangaroo that had gone missing in Oklahoma the year before. It was also not a kangaroo that belonged to a nearby exotic animal farm (all of their kangaroos were accounted for). Eventually, the kangaroo was traced to (and subsequently recaptured by) another guy in the same town who also had a bunch of kangaroos for some reason. So yeah, a kangaroo encounter is probably more likely than you think.

The question raised by this long and sometimes furry tail of unlikely events (including unconstrained extra-Australian macropods) is to what extent autonomous car companies should reasonably be expected to anticipate and devote resources to preparing for things that almost never happen. It seems obvious that an autonomous car (in North America) should know what to do when it senses a deer on the road. You're less likely to encounter a moose: should your car know what to do then? What about a bear? I’m sure kangaroos are farther down the list, but it’s a list that's infinitely long, and there’s no clear point at which a potential situation goes from “worth preparing for” to "not worth the effort.” More to the point, if your autonomous car runs into an occasionally airborne pouched hopper, who is at fault?

We don't mean to suggest that an autonomous car that has not been specifically programmed to avoid a kangaroo is just going to plow straight into the first kangaroo that it sees, since “Don't hit things” is useful, practical advice that tends to be pretty high up on an autonomous car's generalized priority list. However, there are many different ways of not hitting a thing depending on what that thing is, and sometimes that advice can even cause problems if the car senses something that it doesn't understand. A common example might be a plastic bag blowing around in the middle of a highway. Since we humans understand what a plastic bag is and what the consequences would be if we were to collide with it (probably  mild) as opposed to what the consequences would be if we tried not to hit it (possibly serious), we can make the correct decision about what to do. And thanks to having lots of experience driving, and also with generally living in the world, we can rapidly assess unfamiliar objects and make informed decisions about them. Robots of all kinds are notoriously bad at this.

Volvo isn't particularly worried about kangaroos specifically; they're a challenge, but an identifiable and easily definable (and therefore solvable) one. There are other Australian things that autonomous cars need to be prepared for as well, including unsealed reads, unmarked roads, and road trains. But again, the real problem here is that Volvo, and every other autonomous car company, has no idea what might try to cross the road. Whether it's a kangaroo, a human on a pogo stick, or something far stranger, autonomous cars will (somehow) need to be ready for it.

The Conversation (0)