Learning Robust Vehicle Navigation Policy Under Interference and Partial Sensor Failures

Autonomous driving has been a research hotspot for several years. However, most researchers focus on the ideal driving condition, where the sensor data is always accurate and clean. In reality, the urban environment is complex and the sensor data can easily be disturbed, which can lead to failures of existing driving approaches. In this paper, we present a multi-sensor fusion based navigation framework to achieve robust navigation under interference and partial sensor failures, which implicitly learns the reliability of each input information and introduces graph neural networks to reconstruct features for robust navigation policy learning. Experiments on the high-fidelity CARLA platform demonstrate the effectiveness of our approach, especially our strong robustness against various disturbances on both the input point cloud data and the visual images.