2021/11/10 by Yandong Liu, Chengzhong Xu, Liu, Yandong +3 · 1 citation
Computer Science · Engineering · #Advanced Neural Network Applications #Autonomous Vehicle Technology and Safety #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Multimodal Machine Learning Applications #Robotics (cs.RO)
paper · pdf · doi:10.48550/arxiv.2111.06017
openalex publication_date 2021/11/10 · openalex created_date 2022/05/05 · openalex updated_date 2026/07/28
Existing imitation learning methods suffer from low efficiency and\ngeneralization ability when facing the road option problem in an urban\nenvironment. In this paper, we propose a yaw-guided imitation learning method\nto improve the road option performance in an end-to-end autonomous driving\nparadigm in terms of the efficiency of exploiting training samples and\nadaptability to changing environments. Specifically, the yaw information is\nprovided by the trajectory of the navigation map. Our end-to-end architecture,\nYaw-guided Imitation Learning with ResNet34 Attention (YILRatt), integrates the\nResNet34 backbone and attention mechanism to obtain an accurate perception. It\ndoes not need high precision maps and realizes fully end-to-end autonomous\ndriving given the yaw information provided by a consumer-level GPS receiver. By\nanalyzing the attention heat maps, we can reveal some causal relationship\nbetween decision-making and scene perception, where, in particular, failure\ncases are caused by erroneous perception. We collect expert experience in the\nCarla 0.9.11 simulator and improve the benchmark CoRL2017 and NoCrash.\nExperimental results show that YILRatt has a 26.27% higher success rate than\nthe SOTA CILRS. The code, dataset, benchmark and experimental results can be\nfound at https://github.com/Yandong024/Yaw-guided-IL.git\n