Volvo admits its self-driving cars are confused by kangaroos
theguardian.comLots of human drivers are confused by kangaroos and kangaroos are confused by cars especially at night. You see an awful lot dead kangaroos on the side of the road in rural Australia.
As an aside, kangaroos in motion are one of the most efficient and elegant mammals. I love watching them bounding away at speed effortlessly dodging scrub and trees - more like flight than running.
What's the general feeling about taking the meat after a collision? Where I grew up, people would butcher the animal if it's not terribly destroyed. Around here there's a wild hog problem, which are edible but not very tasty, so the local wildlife agency picks them up and they're used in soup kitchens to feed the homeless.
If you hit a kangaroo (I have only ever had a side on collision where the kangaroo jumped into the side of my car and bounced off) you need to get out and check if it has a live joey (baby kangaroo) in its pouch. You are not allowed to keep the joeys, but there are organisations that will take them in and raise them enforce release back into the wild.
Note: If you see a kangaroo (or any marsupial) dead on the side of the road, if it has a pink squiggle spray-painted on its side, someone has already checked it for a joey - so you don't have to try stick your hand into a gooey mess.
Not being Australian: would it be legal?
I grew up in Alaska and when you ran into a moose you weren't allowed to keep it. Instead there was a volunteer crew that was called to clean it up, compensated with the corpse (which is how I grew up on moose).
In France this is illegal for at least two reasons:
- health: most people don’t know how to process raw meat or recognize a sick animal (rabies or other illnesses)
- safety: those animals are tough, it’s not uncommon for boars to wake up from the concussion before, while, or after being loaded in a car, and the last thing you want while driving is a wild, fearful, angry animal suddenly waking up and ripping your car apart from the inside while you’re driving (happens).
- law: hunting is regulated so as to control wild animal population growth, killing some has to be reported one way or another.
I suspect there is another historical reason, akin to why you can be prosecuted for attempting to obliterate your own existence: your own being is basically owned by the Republic (used to be the King), so committing suicide is a prejudice to the State. I can see something like this being in effect: wild animals are a property of the State, and killing them even with your car is poaching. That may be why you’re supposed to turn the corpse to the Gendarmerie (cops branch of the military).
(IANAL, correct me if I’m wrong, that just what I’ve been taught as a kid growing in a rural area full of forests)
It wouldn't be illegal, but its not like hitting a dear or moose in the snow. It's really hot where kangaroos are in Australia, so keeping the meat fresh is the biggest issue.
Also you can get kangaroo meat cheaply at most butchers. It's just like venison.
I have actually heard of people saying they salvage kangaroo (or buffalo) meat after a collision. The tail is considered a delicacy.
Of course, that is assuming you are still OK after hitting one - Kangaroos are BIG critters and can do considerable damage to the vehicle (and occupants) in a collision.
Eating kangaroo is relatively common in Australia. We quite often brag that we are the only country that eats its national coat of arms.
It is quite divisive though, some people enjoy the taste, myself included, but others have a strong dislike. It is high in protein so often enjoyed by gym-going types.
Can't say about kangaroos, but many (did I say many?) years ago I was on a car when we hit a wallaby by accident.
Wallabies are like 1/5 or less in size/weight when compared to kangaroos, they are small animals I would say comparable in size/weight to a large rabbit.
The car was a not so small car, a Holden Torana, a V8 4.2 liters, if I recall correctly, still the hit was very hard and the australian guy that was driving didn't manage to keep the car on the road and we ended up in the field nearby.
No physical harm to anyone on the car, but the fender and front of the car were seriously damaged.
I wouldn't have wanted to have hit a full-sized kangaroo instead of the poor little wallaby.
I met a guy who does it, but just takes the tails. He considered the roo fresh if the blood hadn't congealed by the time he found it.
This was in an environment where daytime temperatures are often 40C (104F) plus during the day.
If it's a fresh kill, yeah if you're really that keen it might be worth cutting the tail off with an axe and cooking it up pretty immediately.
But Roo's need to be bled properly otherwise the meat goes to hell. You can get away with eating fresh tail, but really a roadkill roo wouldn't be very nice.
Plus, after 30 minutes in the sun, it'd be no good for anyone other than the flies.
City dwellers generally wouldn't do so, although I don't know of any laws against it. Some people who grew up on farms or in remote towns certainly do.
Having had plenty of wallabies (smaller version of Kangaroos) jump out in front of me at dawn/dusk, I concur.
The hopping movement makes it incredibly difficult to judge their speed and direction. They actually cover horizontal space at a far greater clip than they seem to, and also means they can make snap changes in their direction of travel VERY quickly, making any avoidance manoeuvre a bit of a wild guess.
They are extremely unpredictable in their change in direction and they will often get panicked by the sound of the car and jump in front rather than away from the car like most other animals.
Unpredictable motion is ideal for evading upright preditors with spears, but no so good for collision avoidance with fast moving metal objects.
Put in another way, unpredictable motion is good for avoiding what's trying to hit you. Not good for avoiding what's trying to miss you.
So, when kangaroo jumps in front of your car your best chance to avoid hit would be to try to steer your car towards it and let kangaroo's instinct do the rest? :)
It's probably complicated by the fright response in kangaroos which is to jump in a random direction. They don't all head in the same direction which is why when you drive past a mob feeding on the verge you often get a couple jumping onto the road in front of you.
I'm kind of worried by the fact that the system judges depth based on visual cues (like a human with one eye closed does) instead of actual depth sensing hardware, to the point that the system needs to recognise kangaroos specifically.
Shouldn't the car stop for any obstacle, even ones it doesn't recognise? They surely can't expect to train it on every possible type of debris.
It doesn't sound like the problem is it's detecting specific animals, but in how it's detecting the distance -using the ground as a reference for things that aren't standing on the ground.
I wonder how it it handles low-flying birds.
When I learned t drive I was instructed not to mind smaller critters as the avoidance maneuver (steering away one way or the other or heavy breaking) is more likely to cause a serious accident than the collision with the small critter.
Unless, of course, you are the small critter.
(When in driving school, I was taught to ignore anything smaller than a reindeer - but assure whoever showed up in distress after I'd put a set of tire marks over their pet's back that I'd tried my best to avoid it.)
True. You also don't want it avoiding a plastic bag that's floating across the road
You needn't swerve but there's no harm in braking.
A risk would be a car suddenly braking for leaves or a plastic bag crossing the road.
What about a plastic sheet obscuring a kangaroo/animal ?
Have you considered a career in QA?
> The company’s “Large Animal Detection system” can identify and avoid deer, elk and caribou, but early testing in Australia shows it cannot adjust to the kangaroo’s unique method of movement.
In Sweden there are elks and deer but not so many kangaroos. :)
An intelligent system needs to learn. To learn it needs examples.
Funny you say that, as I was just reading this: https://www.nrk.no/rogaland/motte-kenguruen-da-han-var-pa-ve... (not Sweden, I know). The video does show their interesting movement.
> "Although it hasn’t been tested in a kangaroo-specific environment, there was an instance where black swans were interfering"
I think I know how to interpret that sentence...
Black Swan Farming (Paul Graham)
Literal black swans are common in parts of Australia. https://en.m.wikipedia.org/wiki/Black_swan
There are no native white swans in Australia. They are all black.
Let's not bring identity politics in to this discussion.
They mean that they have in the past dealt with animal-specific issues, rather than a generic approach. I take it mean that they will have to write some roo-specific rules as they once did for black swans.
Speaking of getting confused by kangaroos...
An old favourite of mine, presumably apocryphal, involving kangaroos and the dangers of reusing code:
Why does it need to recognize specific types of animals? It seems like bad programming practice to code (or learn) for each specific case. Is there not a way to program cars to avoid animals more broadly, or simply objects that are moving towards the road?
I don't think I personally would deal well with a kangaroo appearing near my car due to my lack of specific experience. They move crazy differently from a dog or deer.
Because movement patterns of Kangaroos are too different from other animals?
I believe these systems are based on neural nets, they are trained, not programmed, to recognize animals.
The letters in "confused by kangaroos" rearrange to spell "SOS! Fake, bouncy dragon."
Who wouldn't be confused?
h/t Gary Matthews, elsewhere.
Moose must be hard to detect too I wonder how they tested for them.
Imagine a 700kg (1,500 pound) moose with velvety brown fur suddenly wanting to cross the road at night in front of your car when you are traveling 120km/h (74mph).
That weight is all up high too on spindly legs. It's like running into a high table with 700kg of weigh on it.
Emus have a similar effect - a good amount of mass sitting up quite high on little thin legs. The things punch straight through the windscreen when you hit them.
Emus have the added bonus too of running directly at you when you sound the horn, instead of away as you'd expect. It's quite uncanny, and problematic at a hundred kilometres an hour.
Doesn't this just suggest that computer vision is practically impossible to perfect, and that cameras + LIDAR would be far better for safely driving a car? You need actual distances to objects, not just fuzzy error-prone inferences.
Volvo self-driving cars have multiple LIDARs and radars, in addition to vision.
This stuff is still early days. I don't see how a problem encountered with a research project could possibly indicate that the whole problem is "practically impossible." It would be weird if they didn't have problems like this while building the technology.
The obvious counter-example are humans.
On a more realistic level, stereoscopic vision seems like a pretty decent way to get distances.
Your obvious counter example might need General Intelligence.
>> You need actual distances to objects, not just fuzzy error-prone inferences.
Humans do pretty good with fuzzy notions of where the animal is. Exact distances aren't needed so much as reasonable estimates. If it is probably in the "need to brake zone" then the car brakes. Plus or minus a few meters isn't a big deal so long as you've got margins to work with. As animals are unpredictable, Volvo certainly has wide margins.
Plus if there is anything the human vision system is optimized for it would be seeing and assessing an unknown large mammal.
History (and personal experience) shows pretty clearly that humans are really bad at predicting the movements of kangaroos, though. (I've never been in a car that hit one, but have had some close calls)
The things seem to be able to effortlessly execute 90 degree turns between hops, at speed.
Safety lesson: if you see kangaroos travelling alongside the road you're driving on, slow way down so you can execute an emergency stop, if needed; they're always only one hop from suddenly being right in front of you.
Safety lesson 2: there is always another kangaroo. If you see a kangaroo bound across the road ahead of you, slow way down; there's likely to be another kangaroo a very short distance behind it. (And safety lesson 2 recurses and also applies to that second kangaroo; there is likely a third kangaroo just far enough behind it that you'll think there's no third kangaroo. And so on.)
Another is - don't let your dogs near a kangaroo around a body of water. The kangaroo hops into the water, the dog follows and the kangaroo holds it under and drowns it.
But apparently humans get confused by kangaroos too, based on the number of collisions.
It's not confusion as other Australians here have said they really do not behave like any other animal and appear out of no where at speed. A actual collision with one normally takes out the car and not just the kangaroo as well.
They're hard to see because of their coat, move quickly and erratically and they're pretty much all muscle.
I'm not surprised they get hit as often as they do particularly at night.
Gotta watch out for them kangaroos: http://www.snopes.com/humor/nonsense/kangaroo.asp
I'm not sure what system Volvo has gone for, but most isn't this just a "train more kangaroo data into the neural network" thing to fix?
From the article it seems the hopping it what makes it tricky. Their model is that the distance to the animal is the same as the distance to where their feet touch the ground. When an animal is in the air the feet appear to touch the ground further away due to perspective.
This is the first time I hear about animal detection in self driving cars. Do other brands also do this?
Most demo's only show human interaction.
Edit: and what about flying trash, leaves and so on? Thinking about it there are a lot of object types we interact with. Leaves for example mean nothing to us because they won't harm us. And we know crows will fly away. But pigeons on the other hand..
Footage of Volvo and the kangaroos from a couple of years ago: https://www.youtube.com/watch?v=2YiZ0aHINBI
Those corner cases can be solved by simulation. Off to the sims, to implement kangaroos, deer and people in wheelchairs chasing ducks with a broomstick - just to be thorough.
Just make sure you disable the kangaroos' rocket launchers before giving your big demo... ;)
And therein lies the problem no one talks about: all these autonomous vehicle technologies have only been tested in non-adversarial situations. Machine learning algorithms do not generalize the way humans do, and it has been shown that they’re fairly easy to fool. This is an active area of research at the moment, and no one really knows how to fix this, at least not without deploying an ensemble of networks with different architectures.
Wallaby damned.
So European intelligence is confused by kangaroos, artificial and otherwise.
Reminds me my comment from a month ago...
At least that is an improvement on Tesla being confused by a large truck and hitting it at full speed. https://www.theguardian.com/technology/2016/jun/30/tesla-aut...