No matter their environment, it is important for robots to ensure safety and prevent malfunctions. An important aspect to robotics is therefore collision-free path planning. Conventional methods such as RRT and potential field have difficulty adapting to dynamically changing environments, and have high calculation cost for iterating or searching. In this research, we propose a new path planning method in which the latent space of Conditional Generative Adversarial Networks (cGANs) represents the workspace where the robot avoids obstacles, and a path is planned in this latent space. The selected path is then mapped to a path in joint space. Since the latent space is collision-free, path planning can be performed by connecting the start and the goal robot posture without iteration or searching processes. We confirmed that our single trained model can handle various types of multiple untrained obstacle conditions.