Advanced Deep Learning for Robotics project: sim2real problem in the context of flying vehicles
run git lfs install before working with the rep
conf\config.yaml contains all the shared paramenters.
conf\agent\ contains agent-specific parameters.
You can select the required agent parameters by specifying defaults section (important - only AFTER _self_ to overwrite default parameters)
If you don't want to edit configs to run the script with the specific action - you can specify it from the run cmd:
python baseline.py agent=ddpg
Any key, written to an agent's configuraion, will ovewrite defaults or will be appended to them.
If you want create an agent-specific training configuration, you have to create file inside conf\training\ folder with the same name as the specific agent file, like conf\training\sac2.yaml
If you want to start the training without any randomization, set env.random_type='Fixed' in conf\config.yaml.
To configure parameters of the fixed env, edit the parameters with prefix default_.
To turn on randomization, set env.random_type='Uniform', and tune lower and upper bounds of the parameters with corresponding configs.
To pass the env state (gravity, enable_wind, wind_power, turbulence_power) to an agent, set training.pass_env_parameters=True.
They will be normalized, according to lower and upper bounds, and added to state (see StateInjectorWrapper from EnvironmentRandomizer.py)