Working with the local motorcycle club Iron Order, Tiffany and our Vehicle Operations team set up a day of riding to collect data about how our cars see motorcycles in different types of complex situations.
How do you train a self-driving car to safely ride alongside motorcycles? One of the first steps involves our perception system, which analyzes sensor data and generates a representation of all actors around the vehicle, including motorcyclists and pedestrians. Making sure our cars “see” motorcycles properly is crucial as we train the Aurora Driver. Our machine learning models learn to recognize objects like motorcycles from incoming data from our radar, lidar, and cameras, but they can only do so if they have a good and diverse set of existing examples that are properly labeled. The better data we have, the better our models get at recognition.
Manual data collection is one part of a multi-prong approach that allows us to quickly get data for commonly-seen objects; we also use simulation and collect large amounts of data as we drive on the road.
Recently, we wanted to gather more data about motorcycles in a range of different and complex situations — moving at different speeds, approaching from different lanes, riding together, and more — to improve recognition performance. So we got a group of motorcyclists together for a day of riding and testing.
We sat down with Tiffany Matz from Vehicle Operations, which focuses on our on-road testing, to learn more about this project. She also shared what motivates her to work here and what she does in her free time (hint: COOKIES).
Tell us about your work at Aurora and what led you here. What motivates you and the team?
Tiffany: I’ve been at Aurora for more than two years. I am the Vehicle Operations Training Lead, which involves, as you might have guessed, a lot of training! I train our safety drivers to operate our vehicles and work with our technical teams as we develop the Aurora Driver.
I got into the industry a few years ago out of curiosity and fell in love with technology, research, and development. I feel like I am helping people by contributing to a technology that will make our roads safer.
The fact that Aurora’s various teams are all working towards a common goal — delivering the benefits of self-driving technology safely, quickly, and broadly — keeps everyone motivated. It’s rewarding to see how our team’s testing feedback directly helps improve the performance of the Aurora Driver.
You recently partnered with the local motorcycle club for a day of riding and data collection. What were you trying to learn?
Tiffany: We wanted to collect additional data on what motorcycles look like riding at different speeds and in different situations. This is valuable information and ensures our perception system properly recognizes motorcycles in all sorts of situations.
In this case, the request came directly from the Perception team, who told us, “Any scenario that mimics real-world car/motorcyclist interaction is worth getting data on.”
We were looking for as much information as possible in a variety of circumstances and partnering with the Iron Order motorcycle club allowed us to get a lot of diverse data quickly. For example, we varied:
- positions relative to the car: including motorcycles in adjacent lanes, the same lane, in front of the car, and behind the car
- approaches: including oncoming motorcycles, motorcycles passing the car, and motorcycles riding in front of the car
- speeds: including scenarios where we ran the approaches listed above at various speeds, as well as scenarios where the car is stopped with a motorcycle in front of it
- types of motorcycles: We ended up having an Indian, four Harleys, a KTM sports dirt bike, and a Yamaha cruiser bike.