[1L2-J-11] Robot and real worlds: cognition of objects and environment

[1L2-J-11-01] Design and Evaluation Image Recognition Sub-tasks to Improve End-to-End Learning Model for Self Driving Cars

〇Jing Shi¹, Hao Zhi Li², Toshiyuki Motoyoshi¹, Tadashi Onishi¹, Hiroki Mori³, Tetsuya Ogata^1,4 (1. Department of Intermedia Art and Science, School of Fundamental Science and Engineering, Waseda University, 2. Department of Modern Mechanical Engineering, School of Creative Science and Engineering, Waseda University, 3. Future Robotics Organization, Waseda University , 4. National Institute of Advanced Industrial Science and Technology)

Keywords:Self Driving Cars, Sub-task, End-to-End learning

Sub-task training for a deep neural network can improve main task performance, for example, for self-driving cars and other tasks. However there is no theoretical design principle that how to make sub-tasks suitable for a main task. In order to improve the self-driving task, searching the optimal sub-tasks design is necessary. In this research, we compared multiple combination of sub-tasks sharing a network to generate driving command. In the research of Li et al.2018, a multi-task learning method used two modules, a perception module (extracting semantic segmentation and depth map) for recognition of surrounding circumstances and a driving module for driving operation. Their multi-task method scored higher generalization performance in unknown environment than previous end-to-end self-driving method. In this research, we conducted experiments to improve the generalization ability of their model by modifying sub-task design. As a result, generating semantic segmentation map as sub-task got the best performance for self-driving cars.