In this paper we show how risk-averse reinforcement learning can In this paper we focus on the option hedging problem in a realistic be used to hedge options. We apply a state-of-the-art risk-averse environment where we exploit the power of Reinforcement algorithm: Trust Region Volatility Optimization (TRVO) to a vanilla Learning (RL). In some sense we aim at replicating and hopefully option hedging environment, considering realistic factors such as improving, in an automatic way, the trader's experience of containing discrete time and transaction costs. Realism makes the problem both risk and hedging costs. While there is an extensive twofold: the agent must both minimize volatility and contain transaction literature on both option hedging [13] and reinforcement learning costs, these tasks usually being in competition. We use the [29], there are very few works on the combined topics, the main algorithm to train a sheaf of agents each characterized by a different ones being [4, 5, 11, 15], which we will analyze in Section 5. risk aversion, so to be able to span an efficient frontier on Here we implement a robust tool capable of providing the trader the volatility-p&l space. The results show that the derived hedging with a hedging signal more accurate than the delta hedge, as it strategy not only outperforms the Black & Scholes delta hedge, is optimized in a realistic environment, with discrete time and but is also extremely robust and flexible, as it can efficiently hedge transaction costs. We achieve this result through the use of riskaverse options with different characteristics and work on markets with RL by applying TRVO [2], an algorithm capable of optimizing different behaviors than what was used in training.