Back to work

AWS DeepRacer

AWS DeepRacer

Archived

A reinforcement learning model for autonomous racing that placed 37th of about 3,940 racers in AWS’s October 2022 Open qualifier.

When
2022
My part
Leader of Insper Dynamics’ DeepRacer team.
  • Python
  • Reinforcement learning
  • AWS
A rendered AWS DeepRacer car on a purple background, next to the AWS DeepRacer League logo

37th of about 3,940 racers in the October 2022 Open qualifier, under the name G3mha. That’s the top 1%.

Nearly all of the result came from the reward function rather than the model. The default reward pays the car for staying near the centre line, which produces a car that drives slowly and safely and loses. Paying instead for speed held through a corner exit, with a penalty curve steep enough to keep it on the track, produces a lap that’s quicker and much more prone to failure. Most of the 2022 season was spent finding where between those two a model still finishes.