A repo of gym environments for continuous classic control (ccc) problems. In
particular, this repo offers gym environments where the observation is the state
of the system as given by the dynamics. In addition to these environments,
normalized environments are provided which contain the real state in the info
output and the observation is instead some normalized version of the state.
The main highlights are:
- non normalized observation corresponding directly to the dynamical state
- normalized observation with dynamical state captured in
info['state'] - environment elapsed time captured in
info['time'] - action spaces are continuous
- system parameters (mass, length, etc.) can be specificed
- reset function (to specify initial conditions) can be specified.
The motivation for this is so that gym environments can be used for control problems where the states/observations are not traditionally normalized. The observations in gym environments are normalized since the observation vector is directly passed into neural network models and dealing with normalization within the neural network for each environment is not feasible. However, this also means that gym environments can't be used for control methods that require the non-normalized states/observations.
The normalized environments are directly extended from the non-normalized environments to ensure the same dynamics are used.
If you want to use a different kind of normalization then feel free to extend the original gym environments and output a normalized observation of your own choosing. Similarly, you can extend the gym environments to disregard the given reward shape and implement your own. See gym wrappers for how to do this. (https://alexandervandekleut.github.io/gym-wrappers/)
An optional dependency for rendering multirotor environments:
The rest of the dependencies will come through with the pip install command.
pip install .As you would any other gym environment. See the examples directory for some example code of creating and stepping through the gym env.