================= Safety Guarded RL

Preliminary experiments to evaluate the convergence and safety of RL algorithms when the policy update is guarded by an safety oracle.

Structure

The implementation is basically split into two parts:

• lti_gym: Learning Environment
• guarded_learner: Safe learner based on stable baselines RL agents

The remainder is structured as follows:

• examples: simple examples how to use it
• experiments: all the experiments and evaluation

Usage

.. code-block:: python

import gym

if __name__ == '__main__':
    env = gym.make('lti_gym:simple_grid-v0')

    env.reset()
    for _ in range(1000):
        env.render()
        env.step(env.action_space.sample())  # pick three continous control actions randomly
    env.close()