Imitative Reinforcement Learning Fusing Mask R-CNN Perception Algorithms

Autonomous urban driving navigation is still an open problem and has ample room for improvement in unknown complex environments. This paper proposes end-to-end autonomous approach that combines Conditional Imitation Learning (CIL), Mask R-CNN with DDPG. In the first stage, data acquisition performed by using CARLA, a high-fidelity simulation software. Data collected CARLA used to train network, which object detection segmentation. The segmented images are transformed into backbone of CIL perform supervised (IL). DDPG means Reinforcement further training second shares learned weights from pre-trained model. combination two methods innovative way considering. benefit it possible speed up considerably obtain super-high levels performance beyond humans. We conduct experiments on benchmark driving. final experiments, our algorithm outperforms original MP 30%, 33%, CIRL 10% most difficult tasks, dynamic new environments weather, demonstrating two-stage framework proposed this shows remarkable generalization capability tasks.