Predictive uncertainty estimation using deep learning for soft robot multimodal sensing

The mechanical compliance of soft robots comes at a cost of higher uncertainty in their sensing and perception, which deteriorates the accuracy of predictive models. Predictive uncertainty, which expresses the confidence behind model predictions, is necessary to compensate for the loss of accuracy in soft robot perceptive models. Nevertheless, developing a general framework to capture uncertainties is further challenged by the complex dynamics of soft robots and the difficulties in sensorizing them. In this work, we present a predictive uncertainty estimation framework based on deep learning for soft robot multimodal sensing. We show that the framework can learn to quantify uncertainty and thus is able to express the confidence associated with the predictions during inference. Being data-driven, it is scalable to different types of soft robots and sensor modalities. We demonstrate the framework on a complex multimodal sensing task where a single flex sensor is used to predict the full-body configuration of a soft actuator, as well as the magnitude and location of external contact force. We also discuss how predictive uncertainties are critical to achieve safe learning and model interpretability in soft robotics.