Digital Twin-Driven Navigation of AUVs Through Sensor Space Matching

Detecting and identifying objects on the seabed using autonomous underwater vehicles (AUVs) is a complex task. However, the challenge is no less demanding when it comes to relocating these objects with another AUV to carry out specific tasks. If the task involves certain risks, the sensors on the relocating AUV must be cost-effective and have lower performance compared to those on the vehicle used for the initial detection. This reduction in performance impacts both navigation and perception sensors. To fully make use of the data collected during previous detection and classification missions, it is synthesized into a detailed digital terrain model (DTM), which is further enhanced with information about the seabed type for each cell. The core concept of our navigation approach is to simulate the perception sensor's output at the AUV's estimated location and then estimate the navigation error by comparing this simulated data with the actual sensor data from the AUV's real location. The simulated data can be thought of as originating from a digital twin of the real operating environment. To minimize the data transformation process, the comparison is conducted directly in the sensor data space, rather than converting the data into a geographic coordinate system. The digital twin generates sensor data based on the pose (position and orientation) the AUV believes it has, and the differences between this simulated data and the real sensor data indicates the navigation error. Computationally-efficient simulation has shown to be helpful in underwater operations and previous work has shown that a precise digital terrain model (DTM), created by combining acoustic sensing with photogrammetry techniques, enables the simulation of realistic sonar images. In this study, the low-cost relocation acoustic sensor is a rotating mono-beam profiling sonar operating in the horizontal plane. The sensor data space is a simple 2D plane, with axes defined by the range to the obstacle and its azimuth. To design, develop, and test the algorithm, we draw an analogy with a 2D LiDAR used in land-based environments. The reference DTM is computed from data acquired using a higher-cost 3D LiDAR. To estimate the navigation error between the real position and the digital twin's (numerical simulation) position, we process two 2D LiDAR data sets: one from the real-world sensor and the other from the digital twin. The 2D LiDAR data are kept in their original coordinate space, defined by the beam angle (azimuth) and the range to the closest detection. The algorithm presented in this paper enables the estimation of navigation errors between the two dataset. The computation is repeated at a typical frequency of 5 Hz to 1 Hz, and the results are sent to the control loop to determine the set point. The accuracy of the estimated navigation error is less than 84% for position and 43% for orientation on the first iteration. The algorithm demonstrates strong robustness in directional estimation, achieving perfect accuracy in the sign of the rotational error and over 72% accuracy in the sign of translational errors, demonstrating the effectiveness of the proposed method for relocating an underwater object.