Available Agents

The SLC package includes a collection of robust RL algorithms accompanied by their less stable baselines. These algorithms are designed with non-recurrent MLP actor-critic models, making them well-suited for fully observable RL environments that do not rely on image data, such as the gymnasium Mujoco and stable-gym environments. The implementation follows a modular approach, allowing for seamless adaptation to different types of environments and neural network architectures.

Stable Agents

Important

As explained in the installation section of the documentation, although the opt_type algorithm variable can be used to train on standard gymnasium environments, the stable RL agents require a positive definite cost function to guarantee stability (and robustness). Several custom environments with positive definite cost functions can be found in the stable-gym and ros-gazebo-gym packages. When using the latter, make sure to set positive_reward to True.

The SLC package currently contains the following theoretically stable RL algorithms:

• Lyapunov Actor-Critic (LAC)
• Lyapunov Actor-Twin Critic (LATC)

Unstable Agents

The SLC package currently contains the following (unstable) baseline RL algorithms:

• Soft Actor-Critic