SHRUMS: Sensor Hallucination for Real-time Underwater Motion Planning with a Compact 3D Sonar

TL;DR

SHRUMS tackles robust 3D underwater autonomous navigation under poor visibility by integrating a compact 3D sonar with a novel sensor hallucination concept that synthesizes enhanced depth data from configurable, non-existent sensors. The pipeline combines ICP-based SLAM for local mapping, fast depth-image synthesis from a local map, rapid mesh generation and convex decomposition, and an optimization-based, swept-volume trajectory planner to yield locally optimal, collision-free 3D paths in real time. Its real-world validation with diverse sonar data demonstrates sub-second planning cycles and resilience to tracking losses, highlighting the practical viability of 3D sonar-driven underwater autonomy and opening avenues for field deployments and transfer-learning-enabled improvements.

Abstract

Autonomous navigation in 3D is a fundamental problem for autonomy. Despite major advancements in terrestrial and aerial settings due to improved range sensors including LiDAR, compact sensors with similar capabilities for underwater robots have only recently become available, in the form of 3D sonars. This paper introduces a novel underwater 3D navigation pipeline, called SHRUMS (Sensor Hallucination for Robust Underwater Motion planning with 3D Sonar). To the best of the authors' knowledge, SHRUMS is the first underwater autonomous navigation stack to integrate a 3D sonar. The proposed pipeline exhibits strong robustness while operating in complex 3D environments in spite of extremely poor visibility conditions. To accommodate the intricacies of the novel sensor data stream while achieving real-time locally optimal performance, SHRUMS introduces the concept of hallucinating sensor measurements from non-existent sensors with convenient arbitrary parameters, tailored to application specific requirements. The proposed concepts are validated with real 3D sonar sensor data, utilizing real inputs in challenging settings and local maps constructed in real-time. Field deployments validating the proposed approach in full are planned in the very near future.

Figures (5)

• Figure 1: (a) A submerged container as seen in a video from by an underwater robot's camera. These extremely poor visibility conditions severely limit vision-based methodologies. (b) A range image from the Sonar 3D-15, showing significantly clearer information about the container's geometry. (c) An example range image produced using sensor hallucination within SHRUMS (the proposed pipeline), enhancing the characteristics of the original sensing capacity in terms of resolution, field of view, and sensing completeness. (d) A 3D path produced by SHRUMS to navigate the observed map.
• Figure 2: The SHRUMS pipeline with its main processing stages, including the integration of a 3D Sonar, ICP-based SLAM, sensor hallucination, mesh reconstruction and convex decompositions, path optimization, and, finally, the path follower informing the robot's controls.
• Figure 3: Intricacies of the 3D sonar's sensing output. In (a), a point cloud from the 3D sonar is shown, though there is incomplete coverage of some surfaces, such as the door at the left side. In (b), local consistency is exploited to provide the entire observed scene using KISS-ICP.
• Figure 4: In (a), the sonar output is shown observing pillars from a deployment under a dock, while (b) shows the useless view from the robot's camera and (c) the local map produced by KISS-ICP. Simulated range images are shown in (d) without and (e) after applying a median filter. As it is shown in higher-resolution images, dense range images, required for smooth mesh reconstruction, may require basic processing.
• Figure 5: Qualitative results from our experiments for the Container, Car, and Pillars environment in each row respectively. The first column offers a planning instance for each at the C1 configuration showing the convex polytopes representation and the path with red arrows. Each following column show in yellow the trajectory of the Aqua2 robot, with white being the map as perceived by the state estimation, using C1, C2, and C3 configurations respectively.