Throughout elementary school, I got to see my sister in action on both my district’s middle and high school robotics teams. Inspired by both her and my own interest in tinkering, I found myself following the same path. By now, I have two FRC competition seasons under my belt: Rapid React (2022) and Charged Up (2023).
For my rookie season, scouting was a key responsibility. For every single match played, we have six people assigned to watch an individual team and record their individual performance. We do this to make sure that we choose the best teams for our alliance. Our team has the reputation of having a good scouting game. In the past, we were sometimes even picked for our insights to other teams. All this exposure to scouting got me thinking of sports analytics: what can I find out from the competition data to potentially help my team?
First, I started to look for useful data. The Blue Alliance website hosts the official FIRST Robotics Competition (FRC) data. Within the FIRST community, there has been an ongoing debate for years about which of the two most popular performance metrics is more useful.
The newer metric is the Offensive Power Rating (OPR), which is a way to quantify a team’s contribution to their alliance’s final score. It uses linear algebra to solve an equation set. Often the set of equations would be an overdetermined system. Hence, least square solutions are used. The Blue Alliance Blog has an article explains the details: The Math Behind OPR — An Introduction. The calculated OPR scores for each competition are available on The Blue Alliance website.
The more established metric is called Expected Points Added (EPA). It’s calculated by the website Statbotics. EPA is modeled after Elo rating system which is a well-known method for ranking chess players. Each player is given an initial rating, and then the rating is updated after each game depending on the predicted and actual outcome. Statbotics modified Elo to adapt to FRC games. Here are the details: The EPA Model.
One thing that is different between chess and FRC game is that FRC is a team sport: each game is played between two alliances, red versus blue. Each alliance consists of three teams/robotics on the field. At each competition, there are two stages: qualifying and playoff. The qualifying games are round robin style. At the end of qualifying round, the top ranked teams become team captains. The captains pick first alliance partner (pick1) according to ranked order. Then the second alliance partner (pick2) are also chosen. It is intuitive that the alliance with higher ranked captain has easier time to win. However, I decided to look into whether alliance partner choices (pick1 and pick2) make differences.
Back in fall 2022, I investigated data from the 2022 game Rapid React. Our team competed in four events: two district events, one district championship, and the 2022 world championship. Getting the data off the websites wasn’t as straightforward as I thought. I also had to learn how to manage data with different shapes in R (via the list structure).
In 2023, January and February were busy build season as usual. In March and May, our team won a district event, did well at our other district event and regional championship, and made it to semifinals of our division at the world championship. When I was finalizing my project in June and July, I decided to add in the data from the 2023 game Charged UP since more data is better and the new data has the same structure.
Finally, I presented my project at the Joint Statistical Meetings in Toronto, Canada on August 8th, 2023. First, I did a brief presentation to a large room of professional statisticians who encouraged me with warm applause. Then I was at e-poster answering their questions for 45 minutes. I even received an invitation to the JSM reception of ASA Section on Statistics and data Science Education, and got an honorable mention for my presentations during their executive meeting.
This project has been a wild journey. I started from robotics but ended up in data science.
Comments