This paper argues in favour of using formal methods to ensure safety of deployed stochastic policies learned by robots in unstructured environments. It has been demonstrated that multi-objective learning alone is not sufficient to ensure globally safe behaviours in such robots, whereas learning-specific methods yield deterministic policies which are less flexible or effective in practice. Under certain restrictions on state-space, modelling safety using probabilistic computational tree logic and ensuring such safety via automated repair can overcome these shortcomings. Promising results are obtained on a realistic setup and pros and cons of such method are discussed.