An Observation Feature Study of Robot Imitation Learning for Autonomous Social Navigation on Construction Sites

While social navigation in complex environments is challenging, the design and selection of observation features become more crucial in an imitation learning system. However, the real-world environment is typically too complicated to adequately represent, necessitating the manual design of features. In this paper, we integrate a simulated environment in Unity 3D for manipulating different types and qualities of observation features in a construction site setting. This simulation is interfaced with neural network models trained offline via the robot operating system (ROS), enabling the online testing of imitation learning algorithms in a virtual environment. The experiment results indicate that the distance to construction workers is essential for avoiding collisions while navigating to the targets. Additionally, depth cameras demonstrate remarkable cost-efficiency among the tested sensors. This work provides valuable insights for researchers developing learning from demonstration (LfD) algorithms and seeking to optimize sensor selection, data quality, and state representation for autonomous agents.