Recent breakthroughs in reinforcement learning show that it’s possible to teach an agent to perform a number of difficult tasks entirely unsupervised and exceed human performance. In this paper we apply some of these techniques to the problem of collision avoidance in 2D space. We propose a novel algorithm called KF-learning that improves model accuracy and reduces generalization error with improved sampling. Finally, we train a model that achieves up to 97% accuracy on collision avoidance tasks surpassing human performance.