Testing the Edge: Preparing Autonomous Vehicles For Swooping Birds
This article is part of our Testing the Edge series, describing how Argo AI engineers create structured tests for real-world driving scenarios from the commonplace to the rare “edge-cases.”
How do self-driving cars handle birds flying in front of them?
As part of its normal operation, an autonomous vehicle (AV) uses technology to detect and identify the objects around it, follows and predicts their paths if they are moving, and adjusts its own route in an effort to maintain the safety of its passengers, goods, and all those in the area, while continuing to its destination.
But birds present an especially complex challenge for human drivers and AVs: they’re often small, they change direction rapidly, and they move through the air in front of vehicles, sometimes closely, unlike almost anything else a driver or AV would encounter on a trip. These behaviors taken together can cause confusion or hesitation for human drivers, and could be an issue for AVs. Thankfully engineers at Argo AI, a Pittsburgh-based autonomy products and services company, are designing systems to respond to them.
Argo prepares its AVs to reason about birds as part of its structured testing process. This involves engineers creating realistic props and setting up specific, simulated traffic scenarios on Argo’s closed course test track in Western Pennsylvania.
The AVs drive through the course to see how the Argo Autonomy Platform, the collection of sensors, computer software, and artificial intelligence software that controls the vehicles, performs.
“It’s all part of an endless development, testing, and learning cycle for the Argo Autonomy Platform,” says Thomas McMenamin, who leads Structured Test Operations at Argo.
Flapping Into the Future
One such structured test is designed to simulate a bird swooping into the view of a moving AV. It involves a professional drone operator, some wire, and a decoy duck from a hunting store.
Engineers begin the test by launching the commercially available drone, now fitted with the decoy duck, complete with flapping cloth wings, suspended below by a wire. The real excitement comes when the drone operator flies the drone-duck right at a moving test vehicle.
Argo’s testing philosophy involves pushing the vehicle to ensure it can navigate even extreme situations, and in this case, that means sending the drone-duck diving down at the car from above, flapping in from dead-ahead, or careening in suddenly from either side.
The Argo Autonomy Platform is designed to take appropriate evasive action if it detects an object in its path, explains McMenamin, adding that it can detect and identify a moving object up ahead, such as a human or an animal.
A bird can appear in front of the vehicle much more suddenly, unexpectedly, and erratically than most other things. Pigeons, in particular, are often on the sidewalk and roads in urban centers and can take flight at a moment’s notice.
Clearly, programming the vehicle to brake suddenly or veer sharply for a bird swoop could have significant implications for other road users, and may not be necessary — birds typically only fly in front of moving vehicles for moments, and do so while maintaining their own version of a “safe distance” from the human objects.
The trick for Argo, and any AV developer, then, is to ensure that their AVs drive not only safely, but comfortably and naturalistically around birds. This means striking a delicate balance so an AV isn’t hesitating when it spots birds flying across the road — as it knows they will pass — but also isn’t charging into them.
“It depends on the circumstances,” says McMenamin. “If the AV is pulling out of a parking spot, we may hang tight for a few seconds to let the bird or birds move out of the way. But if the vehicle is driving, we want to be sure that the Autonomy Platform doesn’t behave in an inappropriate way.”
By this, he means the vehicle should not make any sudden or abnormal decisions. “We need to make sure that our riders, other nearby vehicles, and vulnerable road users are safe.”
By carrying out the bird swoop test, the test engineers can see how well the Autonomy Platform tracks the bird. They can then examine the predicted trajectories, and work with the software development team to improve performance. They can also use the test to assess the Autonomy Platform’s height limit, the highest point above the vehicle at which it is able to detect and react to an incoming bird.
Beyond exploring the vehicle’s response to a bird, the test engineers can also test the predicted motion of unknown objects that might be falling out of the sky or floating towards or past the sensors, such as a plastic bag, bugs, or leaves, explains McMenamin. Engineers can also gain an understanding of what might happen if a flock of birds were to fly at the vehicle, he adds.
What might initially seem like an obscure or quirky test, in fact highlights the technical and design creativity required to fine tune a vehicle’s response to real world driving conditions, even if those driving conditions are impacted by things in the sky.