Lateral Ego-Vehicle Control Without Supervision Using Point Clouds

Existing vision based supervised approaches to lateral vehicle control are capable of directly mapping RGB images the appropriate steering commands. However, they prone suffering from inadequate robustness in real world scenarios due a lack failure cases training data. In this paper, framework for more robust and scalable model is proposed. The only requires an unlabeled sequence images. trained takes point cloud as input predicts offset subsequent frame which angle inferred. poses turn obtained visual odometry. conceived by projecting dense depth maps into 3D. An arbitrary number additional trajectories can be generated during training. This increase model. Online experiments conducted on driving simulator show that performance our superior same initial data set comparable but collected with noise injection.