UnityInvertedPendulumTraining

Dependency setup

• Create a venv for the project: python3 -m venv ~/python-envs/robot-spin-env
• Activate: source ~/python-envs/robot-spin-env/bin/activate
• Install ML-Agents: pip3 install mlagents

Run training

mlagents-learn config/trainer_config.yaml --run-id="<somerunid>"

Run inferenece

• mlagents-learn config/trainer_config.yaml --run-id="<somerunid>" --inference
• Not really a point to doing this over just running with a brain in the editor itself.

Run training with curriculum

mlagents-learn config/trainer_config.yaml --run-id="<somerunid>" -train --curriculum=config/curriculum.yaml

View results in tensorboard

• tensorboard --logdir=summaries
• Navigate to localhost:6006

Debugging

• For incompatability between apis (message like "The communication API version is not compatible between Unity and python") try checking out the latest branch of ML-Agents then reinstalling with:
  • cd ml-agents-envs && pip install -e . && cd ../ml-agents && pip install -e .
  • Then you may have to relaunch your venv to see the changes

Other notes

• Not sure if ml-agents and ml-agents-envs are required, but gave up on debugging.