Clarifications about the Deepracer Challenge

Hi Everyone,

I’ll clarify some things about the competition that some participants have raised.

  1. What is the competition about?

    The primary goal of the competition is Simulation to Real world transfer. The simulator is provided to participants to familiarize and train on the environment provided, Round 1 serves this purpose because the real track is only in Round 2. To that end, we have not provided any settings to participants to change the environment or access internal states of the environment like the which won’t be available in the real world track.

  2. How do I modify the reward function

    You can modify the reward function by using the observations directly. However, we understand that having access to internal state variables will help with this, so we are working on releasing them. However, we suggest to only use them for the reward function as they won’t be available during inference.

  3. Are the evaluation environments seeded

    Yes, evaluation environments are seeded. However, we emphasize that the real world track is the main focus, the simulation leaderboard scores will be used to filter low scoring submissions. If any discrepancies come up for the evaluation scores feel free to contact us, we will rerun the submissions to keep things fair.

  4. Facing issues running docker

    The current simulator is a somewhat heavy in compute usage, so some participants are facing issues running it. We are working on making it more user friendly and will release changes soon.