Attention

This challenge has ended!

This documentation is only for the Real Robot Challenge 2020 which has ended. Following challenges have their own documentation, see the challenge website for more information.

Example Training and SubmissionΒΆ

In the folder rrc_simulation/example you can find a full example of a training method and how to create a submission using the trained policy:

  • example_pushing_training_env.py is a gym environment adapted from the standard cube_env.py. It is adapted to facilitate training. Note however, that care has been taken to not modify the simulation itself, such that state-action trajectories remain coherent with cube_env.py.

  • train_pushing_ppo.py trains a pushing policy on the env defined above, using PPO.

  • view_pushing_ppo.py loads a trained policy and visualizes it. For convenience, the repository already contains some trained policies in the folder training_checkpoints. You may for instance visualize a policy by executing:

    python view_pushing_ppo.py --model_path ./training_checkpoints/ --time_steps 78000000

  • rrc_simulation/example/evaluate_policy.py was adapted from rrc_simulation/scripts/evaluate_policy.py. This file allows you to run the full evaluation on the trained policy by executing:

    rrc_evaluate path/to/output_directory

    in the rrc_simulation/example folder. Note that on our side, we will run rrc_evaluate at the default location of the evaluate_policy.py which is the scripts folder. So if you wanted to use this example as an actual submission, you would replace rrc_simulation/scripts/evaluate_policy.py with rrc_simulation/example/evaluate_policy.py.