After I returned from Germany, I worked as a vehicle operator at another self-driving vehicle company before making my way to Aurora. I decided to join Aurora because I wanted to stay in Pittsburgh, and I saw Aurora as the most promising self-driving vehicle company in the area.
What does the Capability Operations team do?
Luke: Capability Operations is a relatively new team, and our job is to supply software engineering groups with high-quality manual driving data for various capabilities. A capability could be something like nudging, which is when the Aurora Driver adjusts its trajectory to get around obstacles in the road.
How do you get good manual driving data, and how does it help improve the Aurora Driver?
Luke: First, we sync with the autonomy teams to figure out exactly what kinds of data they need. For example, this year we were asked to gather examples of nudging around oncoming vehicles breaching our lane, nudging around double-parked vehicles on the road-side, etc.
Then, we put together routes for our trained vehicle operators to drive, where they can successfully demonstrate a maneuver many times in different circumstances. Finally, we review that incoming data to ensure it meets our quality standards, tag it with the maneuver type and a time stamp, and send it along to engineering.
That driving data is fed into machine learning models that teach the Aurora Driver how to do that maneuver smoothly. This is incredibly useful because it helps the Aurora Driver make decisions that feel natural on the road. We’re building a Driver that follows road rules, while also capturing that feeling of how trained vehicle operators actually drive.
Can you walk us through an example of a log you’d send to the motion planning team?
Luke: Below, our vehicle operators manually performed a smooth unprotected left turn. This example helped teach the Aurora Driver how to find an appropriate gap in oncoming traffic. It also demonstrated a situation when it was appropriate to pull up past the stop line at a traffic light.