Trust-Aware Dynamic Navigation for Mobile Robots with Sensor Noise

Abstract

Mobile robots are increasingly using learning-based methods to navigate in dynamically changing environments. Even though achieving impressive task performance in practice, these methods at best do not take into account uncertainty on sensory observations and model predictions, which can result in unsafe or undesired behavior. In this paper, we introduce a confidence-based adaptive navigation system in which the influence of learned control is dynamically tempered by the confidence of the prediction. The framework combines uncertainty-aware learning and reactive safety controller. Passive estimation of confidence monitors the trustworthiness of learned navigation policies and adjusts control authority. Confident, adaptive control wins out when confidence is high and confidence degradation results in reactive safety behaviors. We evaluate the approach via simulation experiments in a Gazebo setup, varying sensor noise. The experimental results show that the proposed approach can greatly enhance navigation robustness, decrease collision possibility, and well preserve real-time tracking comparing with traditional learning-based or reactive methods. This confirms the significance of confidence-aware control for secure and efficient mobile robot navigation.