We present a novel acoustic camera simulator (ACSim) that generates realistic sonar images by incorporating recursive ray tracing and sonar artifact modeling and provides various ground truth labels, enabling benchmarking and learning purposes. The 2D forward-looking sonar (FLS), also known as the acoustic camera, produces high-quality 2D images. Conducting real-world underwater experiments is challenging, making realistic sonar image simulation a necessary alternative. However, existing simulators often lack sufficient realism or are limited to specific scenes and phenomena. As a result, training on simulations and testing on real sonar images (i.e., sim-to-real) remains an open problem for deep learning-based applications. Our work introduces a novel sonar simulator with a customized rendering engine. We use recursive ray tracing to model multipath reflections in arbitrary scenes and propose physics-based shading for intensity computation. We propose a resampling method for anti-aliasing and model significant artifacts such as rolling shutter distortions and cross-talk noise. The simulator provides various ground truths for benchmarking and deep learning applications. We tested several tasks by training on synthetic images and demonstrated that the models also work on real images. We developed a Blender add-on for an enhanced user interface and will make the simulator open-source to advance future research.

2D forward-looking sonar is a crucial sensor for underwater robotic perception. A well-known problem in this field is estimating missing information in the elevation direction during sonar imaging. There are demands to estimate 3D information per image for 3D mapping and robot navigation during fly-through missions. Recent learning-based methods have demonstrated their strengths, but there are still drawbacks. Supervised learning methods have achieved high-quality results but may require further efforts to acquire 3D ground-truth labels. The existing self-supervised method requires pretraining using synthetic images with 3D supervision. This study aims to realize stable self-supervised learning of elevation angle estimation without pretraining using synthetic images. Failures during self-supervised learning may be caused by motion degeneracy problems. We first analyze the motion field of 2D forward-looking sonar, which is related to the main supervision signal. We utilize a modern learning framework and prove that if the training dataset is built with effective motions, the network can be trained in a self-supervised manner without the knowledge of synthetic data. Both simulation and real experiments validate the proposed method.

Mobile robot navigation in a human-populated environment has been of great interest to the research community in recent years, referred to as crowd navigation. Currently, offline reinforcement learning (RL)-based method has been introduced to this domain, for its ability to alleviate the sim2real gap brought by online RL which relies on simulators to execute training, and its scalability to use the same dataset to train for differently customized rewards. However, the performance of the navigation policy suffered from the distributional shift between the training data and the input during deployment, since when it gets an input out of the training data distribution, the learned policy has the risk of choosing an erroneous action that leads to catastrophic failure such as colliding with a human. To realize risk sensitivity and improve the safety of the offline RL agent during deployment, this work proposes a multipolicy control framework that combines offline RL navigation policy with a risk detector and a force-based risk-avoiding policy. In particular, a Lyapunov density model is learned using the latent feature of the offline RL policy and works as a risk detector to switch the control to the risk-avoiding policy when the robot has a tendency to go out of the area supported by the training data. Experimental results showed that the proposed method was able to learn navigation in a crowded scene from the offline trajectory dataset and the risk detector substantially reduces the collision rate of the vanilla offline RL agent while maintaining the navigation efficiency outperforming the state-of-the-art methods.

Retrieving the missing dimension information in acoustic images from 2D forward-looking sonar is a well-known problem in the field of underwater robotics. There are works attempting to retrieve 3D information from a single image which allows the robot to generate 3D maps with fly-through motion. However, owing to the unique image formulation principle, estimating 3D information from a single image faces severe ambiguity problems. Classical methods of multi-view stereo can avoid the ambiguity problems, but may require a large number of viewpoints to generate an accurate model. In this work, we propose a novel learning-based multi-view stereo method to estimate 3D information. To better utilize the information from multiple frames, an elevation plane sweeping method is proposed to generate the depth-azimuth-elevation cost volume. The volume after regularization can be considered as a probabilistic volumetric representation of the target. Instead of performing regression on the elevation angles, we use pseudo front depth from the cost volume to represent the 3D information which can avoid the 2D-3D problem in acoustic imaging. High-accuracy results can be generated with only two or three images. Synthetic datasets were generated to simulate various underwater targets. We also built the first real dataset with accurate ground truth in a large scale water tank. Experimental results demonstrate the superiority of our method, compared to other state-of-the-art methods.