Assistive robots introduce a new paradigm for developing advanced personalized services. At the same time, the variability and stochasticity of environments, hardware and unknown parameters of the interaction complicates their modelling , as in the case of staircase traversal. For this task, we propose to treat the problem of robot configuration control within a reinforcement learning framework, using policy gradient optimization. In particular, we examine the use of safety or traction measures as a means for endowing the learned policy with desired properties. Using the proposed framework, we present extensive qualitative and quantitative results where a simulated robot learns to negotiate staircases of variable size, while being subjected to different levels of sensing noise.