The aim of the thesis is to implement and validate a new Actor-Critic algorithm for safe reinforcement learning based on Pareto duality. The student will develop a regularized Actor-Critic agent where the critic learns a dual Q-function and the actor is trained to match a target policy that balances performance, stability, and exploration. The final implementation will be evaluated against a state-of-the-art method on appropriate safe reinforcement learning environments. The text part of the thesis will contain a high-level description of the implemented algorithm, description of benchmarks, and thorough presentation and discussion of the experimental results.