In Scientific reports ; h5-index 158.0
Semi-structured environments are difficult for autonomous driving because there are numerous unknown obstacles in drivable area without lanes, and its width and curvature considerably change. In such environments, searching for a path on a real-time is difficult, and localization data are inaccurate, reducing path tracking accuracy. Instead, alternative methods that reactively avoid obstacles in real-time using candidate paths or an artificial potential field have been studied. However, these require heuristics to identify specific parameters for handling various environments and are vulnerable to inaccurate input data. To address these limitations, this study proposes a method in which a vehicle drives toward drivable area using vision and deep learning. The proposed imitation learning method learns the look-ahead point where the vehicle should reach on a vision-based occupancy grid map to obtain a safe policy with a clear state action pattern relationship. Furthermore, using this point, the data aggregation (DAgger) algorithm with weighted loss function is proposed, which imitates expert behavior more accurately, especially in unsafe or near-collision situations. Experimental results in actual semi-structured environments demonstrated the limitations of general model-based methods and the effectiveness of the proposed imitation learning method. Moreover, simulation experiments showed that DAgger with the weight obtains a safer policy than existing DAgger algorithms.
Ahn Joonwoo, Kim Minsoo, Park Jaeheung
2022-Dec-09
