Humanoid robots are revolutionizing the landscape of robotic research. The current availability of cheap humanoid robots is opening new possibilities in a wide variety of industrial and healthcare applications. Indeed, the human form factor allows these robots to integrate easily into environments designed for humans, allowing them to work both in tight spaces and crowded areas. Furthermore, unlike mobile platforms, humanoids can overcome obstacles (such as cables) and climb stairs. So far, the humanoid revolution has been powered by Reinforcement learning technology, which allows learning of robust and agile locomotion policies. However, Reinforcement Learning techniques are not enough to solve more advanced tasks of Humanoid robots: locomanipulation. This problem is much more challenging as it involves contacts. The problem becomes more challenging if we need to ensure the safety and robustness of the policy in safety-critical environments, where a mistake could damage valuable goods or harm people. This project aims to face these issues by combining Reinforcement Learning with safety techniques for the loco-manipulation task. Building from the experience of the main applicant in Safe Reinforcement Learning and locomotion, this project will allow us to deploy Reinforcement Learning policies to solve specific loco-manipulation tasks, opening the possibilities of deploying humanoid robots in advanced industrial settings or in the vicinity of humans.