Environment package

This package contains some useful wrappers, as well as termination and rewards functions that can be passed to mbrl.models.ModelEnv to use for simulation, if needed. One useful class is mbrl.env.mujoco_envs.MujocoGymPixelWrapper, which facilitates the use of pixel-based observations in mujoco-gym environments.

Additionally, for completion, we provide a number of custom environments that have been used in the original papers of the algorithms we have implemented. The environments are provided to match their original versions, with the only changes related to the Mujoco version (we only support 2.0). Whenever possible, we suggest to use the standard implementations of these environments in libraries like ``mujoco-gym``, ``dmcontrol``, rather than using the environments in this folder. These are mostly provided for completeness, and to facilitate debugging of the algorithms provided. The current custom environments are:

• cartpole_continuous: a basic continuous version of gym’s cartpole environment. The only change is that the force applied is multiplied by an action in the range (-1, 1).

• ant_truncated_obs` and ``humanoid_truncated_obs: these are the versions of Ant-v2 and Humanoid-v2 used in the original MBPO paper, which removes some dimensions from the full observation.

• pets_halfcheetah, pets_cartpole, pets_reacher, and pets_pusher are the environments used in the original PETS paper, which include some observation pre-processing functions.