Hands-on Sessions & Challenge

The MRSS 2023 hands-on sessions and competition will explore how to build an autonomous visual navigation pipeline for the Unitree Go1/Go2 Quadruped Robot. Participants will get the opportunity to learn about the theoretical and practical aspects of Simultaneous localization and mapping (SLAM). We will use visual fiducial tags (April Tags) for mapping an obstacle course and navigate to specified goal locations.

For the final challenge, participants will be expected to investigate how Deep Reinforcement Learning (DRL) can contribute to the control and/or planning aspects of the navigation pipeline. DRL policies will be trained in the Isaac Gym simulator and deployed on a Go1 unit to solve navigation tasks involving previously unseen obstacle layouts.

Navigation task

There are six teams and two arenas. Each team will get 10 minutes to evaluate their method on the hardware three times in each arena.

1. Arenas:
   1. We will set up a set of obstacle courses for the robot to traverse.
   2. We will adjust these obstacle courses such that the layout changes before the final compeition. This way, we can’t use hard-coded navigation systems that don’t adapt to the changes in the arena.
      1. The arenas will include large obstacles with marked tags that should be avoided to succeed.
      2. Additionally, low-terrain obstacles, will be added to the map that the robot will have to traverse
2. Evaluation:
   1. Teams will be scored based on their distance from the goal when their robot stops moving.
   2. THE LOWEST SCORE IS THE BEST
   3. +25 points for running into the goal.
   4. +50 pts for running into the obstacle
   5. Each team will get three chances to evaluate their solution on two different arena layouts.