Robust Bayesian Neural Network Algorithms for Reliable Uncertainty Quantification in Robotics

Abstract

Accounting for uncertainties while creating a model is useful in designing robots for environments that contain uncertainty due to noise, dynamics, or partial observability. In this case, a Bayesian Neural Network can become a potential solution. This article provides a description of the BNN with computationally efficient variational inference with weight Gaussian regularization, producing calibrated uncertainty estimates. It is demonstrated how effective this algorithm is by using it on robot navigation problems based on the TurtleBot2 Navigation Dataset and robot manipulation on the PyBullet Robotics Suite. Measures of effectiveness considered are prediction accuracy, expected calibration error (ECE), negative log-likelihood (NLL), and inference speed. Based on the results, it can be concluded that the proposed algorithm works sufficiently well, providing a prediction accuracy of 90.5%, the lowest ECE, i.e., ECE = 0.04, and computational efficiency close to that of neural networks, MC dropout, deep ensembles, which are state-of-the-art approaches. Thus, the effectiveness of the proposed algorithm in robot applications is proven. Potential applications include autonomous navigation, sensor fusion, and manipulation, making robots more adaptable and safer. Future research may involve reinforcement learning, multi-agent systems, and sensory inputs.