Due to complex oceanic environments, underwater gliders typically must satisfy a variety of environmental conditions in order to complete high level objectives. Underwater navigation, for example, requires that a glider must periodically surface and re-localize in order to ensure adequate progress is being made. Such conditions may be directly encoded in Hierarchical Task Network (HTN) planners to ensure that glider actions are valid over the execution of a plan. However, HTN planners may not be able to find good solutions when actions have uncertain costs, such as when a glider is disturbed by a flow field. We propose a bounded cost HTN planner that leverages a modified potential search method in order to find good navigation plans that satisfy user-defined constraints. Simulation results are presented to validate the approach.
more »
« less
Monocular Camera and Single-Beam Sonar-Based Underwater Collision-Free Navigation with Domain Randomization
Underwater navigation presents several challenges, including unstructured unknown environments, lack of reliable localization systems (e.g., GPS), and poor visibility. Furthermore, good-quality obstacle detection sensors for underwater robots are scant and costly; and many sensors like RGB-D cameras and LiDAR only work in-air. To enable reliable mapless underwater navigation despite these challenges, we propose a low-cost end-to-end navigation system, based on a monocular camera and a fixed single-beam echo-sounder, that efficiently navigates an underwater robot to waypoints while avoiding nearby obstacles. Our proposed method is based on Proximal Policy Optimization (PPO), which takes as input current relative goal information, estimated depth images, echo-sounder readings, and previous executed actions, and outputs 3D robot actions in a normalized scale. End-to-end training was done in simulation, where we adopted domain randomization (varying underwater conditions and visibility) to learn a robust policy against noise and changes in visibility conditions. The experiments in simulation and real-world demonstrated that our proposed method is successful and resilient in navigating a low-cost underwater robot in unknown underwater environments. The implementation is made publicly available at https://github.com/dartmouthrobotics/deeprl-uw-robot-navigation.
more »
« less
- NSF-PAR ID:
- 10400860
- Editor(s):
- Billard, A.; Asfour, T.; Khatib, O.
- Date Published:
- Journal Name:
- International Symposium on Robotics Research (ISRR)
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
null (Ed.)Abstract This work proposes vision-only navigation strategies for an autonomous underwater robot. This approach is a step towards solving the coverage path planning problem in a 3-D environment for surveying underwater structures. Given the challenging conditions of the underwater domain, it is very complicated to obtain accurate state estimates reliably. Consequently, it is a great challenge to extend known path planning or coverage techniques developed for aerial or ground robot controls. In this work, we are investigating a navigation strategy utilizing only vision to assist in covering a complex underwater structure. We propose to use a navigation strategy akin to what a human diver will execute when circumnavigating around a region of interest, in particular when collecting data from a shipwreck. The focus of this article is a step towards enabling the autonomous operation of lightweight robots near underwater wrecks in order to collect data for creating photo-realistic maps and volumetric 3-D models while at the same time avoiding collisions. The proposed method uses convolutional neural networks to learn the control commands based on the visual input. We have demonstrated the feasibility of using a system based only on vision to learn specific strategies of navigation with 80% accuracy on the prediction of control command changes. Experimental results and a detailed overview of the proposed method are discussed.more » « less
-
Vision-based state estimation is challenging in underwater environments due to color attenuation, low visibility and floating particulates. All visual-inertial estimators are prone to failure due to degradation in image quality. However, underwater robots are required to keep track of their pose during field deployments. We propose robust estimator fusing the robot's dynamic and kinematic model with proprioceptive sensors to propagate the pose whenever visual-inertial odometry (VIO) fails. To detect the VIO failures, health tracking is used, which enables switching between pose estimates from VIO and a kinematic estimator. Loop closure implemented on weighted posegraph for global trajectory optimization. Experimental results from an Aqua2 Autonomous Underwater Vehicle field deployments demonstrates the robustness of our approach over different underwater environments such as over shipwrecks and coral reefs. The proposed hybrid approach is robust to VIO failures producing consistent trajectories even in harsh conditions.more » « less
-
Bearing only cooperative localization has been used successfully on aerial and ground vehicles. In this paper we present an extension of the approach to the underwater domain. The focus is on adapting the technique to handle the challenging visibility conditions underwater. Furthermore, data from inertial, magnetic, and depth sensors are utilized to improve the robustness of the estimation. In addition to robotic applications, the presented technique can be used for cave mapping and for marine archeology surveying, both by human divers. Experimental results from different environments, including a fresh water, low visibility, lake in South Carolina; a cavern in Florida; and coral reefs in Barbados during the day and during the night, validate the robustness and the accuracy of the proposed approach.more » « less
-
Semantic cues and statistical regularities in real-world environment layouts can improve efficiency for navigation in novel environments. This paper learns and leverages such semantic cues for navigating to objects of interest in novel environments, by simply watching YouTube videos. This is challenging because YouTube videos do not come with labels for actions or goals, and may not even showcase optimal behavior. Our method tackles these challenges through the use of Q-learning on pseudo-labeled transition quadruples (image, action, next image, reward). We show that such off-policy Q-learning from passive data is able to learn meaningful semantic cues for navigation. These cues, when used in a hierarchical navigation policy, lead to improved efficiency at the ObjectGoal task in visually realistic simulations. We observe a relative improvement of 15-83% over end-to-end RL, behavior cloning, and classical methods, while using minimal direct interaction.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- The DOI is not currently available.
