 This paper proposes a novel approach for designing a robotic orthosis controller that considers physical human-robot interaction, PHRI. First, a deep reinforcement learning, deep RL, based imitation learning algorithm is used to generate a human-gate model from a healthy subject. Then, a human model with a weakened soleus muscle is created by modifying the original model. Finally, a robotic orthosis is attached to the right ankle of the modified model, and a deep RL-based policy is trained to provide assistance during walking. The proposed approach was validated through comparison of simulation results with experimental data, as well as application of the learned policy to ankle orthoses. This article was authored by Jang In-haon, Jang Hoon Lee, Hosee and Choi, and others.