Abstract— Reinforcement learning (RL) is a powerful tool for

training agents to perform complex tasks. However, from time-to-
time RL agents often learn to behave in unsafe or unintended

ways. This is especially true during the exploration phase, when
the agent is trying to learn about its environment. This research
acquires safe exploration methods from the field of robotics and
evaluates their effectiveness compared to other algorithms that are
commonly used in complex videogame environments without safe
exploration. We also propose a method for hand-crafting
catastrophic states, which are states that are known to be unsafe
for the agent to visit. Our results show that our method and our
hand-crafted safety constraints outperform state-of-the-art
algorithms on relatively certain iterations. This means that our
method is able to learn to behave safely while still achieving good
performance. These results have implications for the future
development of human-level safe learning with combination of
model-based RL using complex videogame environments. By
developing safe exploration methods, we can help to ensure that
RL agents can be used in a variety of real-world applications, such
as self-driving cars and robotics.