High-Speed Collision Avoidance using Deep Reinforcement Learning and Domain Randomization for Autonomous Vehicles

Recently, deep neural networks trained with Imitation-Learning techniques have managed to successfully control autonomous cars in a variety of urban and highway environments. One of the main limitations of policies trained with imitation learning that has become apparent, however, is that they show poor performance when having to deal with extreme situations at test time - like high-speed collision avoidance - since there is not enough data available from such rare cases during training. In our work, we take the stance that training complex active safety systems for vehicles should be performed in simulation and the transfer of the learned driving policy to the real vehicle should be performed utilizing simulation-to-reality transfer techniques. To communicate this idea, we setup a high-speed collision avoidance scenario in simulation and train the safety system with Reinforcement Learning. We utilize Domain Randomization to enable simulation-to-reality transfer. Here, the policy is not trained on a single version of the setup but on several variations of the problem, each with different parameters. Our experiments show that the resulting policy is able to generalize much better to different values for the vehicle speed and distance from the obstacle compared to policies trained in the non-randomized version of the setup.