Dynamic Interval Restrictions on Action Spaces in Deep Reinforcement Learning for Obstacle Avoidance