Google has taught its self-driving cars to recognize cyclists and how they behave to better predict their course.
The self-driving car, using sensors and software, can also detect and interpret cyclists’ hand signals, Google says in a report released Tuesday. The self-driving cars have learned to be more cautious around cyclists as a result. The cars have also learned how to recognize different types of bikes such as tandems and unicycles.
From Google:
Our cars recognize cyclists as unique users of the road, and are taught to drive conservatively around them (it helps to have a number of avid cyclists on our engineering team!).
Through observing cyclists on the roads and private test track, we’ve taught our software to recognize some common riding behaviors, helping our car better predict a cyclist’s course. Our sensors can detect a cyclist’s hand signals as an indication of an intention to make a turn or shift over. Cyclists often make hand signals far in advance of a turn, and our software is designed to remember previous signals from a rider so it can better anticipate a rider’s turn down the road. Because our cars can see 360 degrees, we’re more aware of cyclists on the road—even in the dark.
Google introduced the prototype—a gumdrop-shaped vehicle it designed itself—in June 2015. The self-driving car doesn’t have pedals or a steering wheel, but only sensors and software. It hopes to commercialize its technology by 2020. The company’s tests still include Lexus RX450h SUVs equipped with autonomous software. Google is testing its self-driving cars—which are equipped with a backup steering wheel and brakes when on public roads—in Mountain View, Calif., Austin, Phoenix, and Kirkland, Wash.
Get Data Sheet, Fortune’s technology newsletter
If Google hopes to launch its self-driving cars for public use in just four years, its software will have to learn all of the nuances required for driving. For instance, the company recently shared how its engineers taught its autonomous prototypes when it’s appropriate to honk, and to use a different honks and beeps depending on the circumstance.
Of course, it needs to be able to see the cyclists too. Google relies on sensors as well as light-sensitive radar (known as Lidar), which are large and expensive. Tesla’s autopilot technology, which allows for hands-free driving on highways, does not use Lidar. Tesla CEO Elon Musk has previously called it unnecessary and overkill for what autopilot needs. But that could change.
Concerns surrounding the safety of autonomous vehicle technology—and the continued fallout over the death of a Tesla driver using autopilot—could prompt Tesla and other companies to use Lidar as autonomous features continue to increase in cars.
Exactly why solely vision based systems don't cut it for driverless cars. Need long range precision depth sensing! https://t.co/KQs0yror7G
— Tony Fadell (@tfadell) July 2, 2016
To be clear, Tesla does use radar, just not Lidar. In later tweets Tony Fadell, the former CEO and co-founder of Alphabet-owned Nest and now co-founder of Actev Motors, say radar or Lidar should be used.
Lidar is used by other automakers that are testing self-driving cars, including Ford. But it’s unclear if the price will fall enough to be deployed widely in commercialized self-driving cars.
