Research

"Research is to see what everybody else has seen, and to think what nobody else has thought." - Albert Szent-Györgyi

For our team, research has been a fundamental part of our growth and progress. We are very proud that the team members have developed these investigations which have helped us solve problems and find solutions. We are very happy that they succeeded and that we can share our findings and knowledge with the world. Enjoy the reading!

This paper presents a low-level controller for an unmanned surface vehicle based on adaptive dynamic programming and deep reinforcement learning. This approach uses a single deep neural network capable of self-learning a policy, and driving the surge speed and yaw dynamics of a vessel. A simulation of the vehicle mathematical model was used to train the neural network with the model-based backpropagation through time algorithm, capable of dealing with continuous action-spaces. The path-following control scenario is additionally addressed by combining the proposed low-level controller and a line-of-sight based guidance law with time-varying look-ahead distance. Simulation and real-world experimental results are presented to validate the control capabilities of the proposed approach and contribute to the diversity of validated applications of adaptive dynamic programming based control strategies. Results show the controller is capable of self-learning the policy to drive the surge speed and yaw dynamics, and has an improved performance in comparison to a standard controller.
This article addresses a guidance and control strategy for an unmanned surface vehicle subject to uncertainties. The controller uses a dynamic three-degree-of-freedom model with identified parameters of an experimental platform. A strategy based on adaptive sliding-mode control is designed for the motion control of surge speed and heading variables, whereas a line-of-sight guidance law, with time-varying look-ahead distance, is applied to achieve path following. The adaptive controller is robust against bounded uncertainties/disturbances, and it does not overestimate the control gain. Outdoor experimental results validate the performance and advantages of the proposed control scheme in three different scenarios: 1) set-point regulation; 2) tracking of time-varying references; and 3) path following. To further demonstrate the characteristics of the proposed controller, a comparison against a standard controller while carrying an extra payload of 20% of the mass of the vessel is included.
This paper presents a guidance and control scheme for an unmanned surface vehicle. The approach combines a deep reinforcement learning based guidance law that can learn the dynamics of a vessel with an adaptive sliding mode controller to achieve path-following. The guidance implements a deep deterministic policy gradient algorithm to obtain the desired heading command, whereas the adaptive control drives the heading and surge speed. The proposed guidance has self-learning ability based on evaluative feedback, which does not require any prior knowledge of the dynamic system, and the controller exhibits robustness against bounded uncertainties and perturbations, control gain non-overestimation, and chattering reduction. Simulation results show that the proposed guidance and control law achieves fast convergence and small overshoot, and improved performance when compared against line-of-sight based guidance laws.
The overall strategy, vehicle design elements, and experimental result from VantTec for RoboSub 2020 are presented. The strategy was constructed as a two-year plan, and from experience from RoboBoat competitions. The higher priority was given to have a robust and reliable hardware and software system to use as a base for future editions, and to develop autonomy capabilities such as object and sound detection and localization, and path-following control. Hence, the Gate and Buoys challenges are the priority tasks to tackle, as they do not require any additional hardware components than the base system. Members were divided in mechanical, electrical, and software areas, with a balance between software or hardware experience. Moreover, the main mechanical and electrical design elements are the frame, enclosures, PCBs, and sensors and actuators, which bring the desired ro-bustness and reliability. Likewise, the system architecture design connects the required components for hardware and software communications, achieving determinism and task prioritization. The software architecture design allows for modular integration of the required software components. Next, computer vision and path-following control algorithms allow the VTec U-III unmanned underwater vehicle to perceive obstacles and navigate through planned waypoints. Furthermore, as the COVID-19 pandemic grew, the proposed strategy adapted to exchange manufacturing time for design time, and all of the challenge-oriented hardware components were designed. Similarly, the course approach was fully implemented in simulation, where results show successful completion of path-following control , and the Gate, Buoys, and, additionally, Torpedo competition tasks. Finally, results indicate the computer vision algorithm is capable of detecting the proposed RoboSub 2020 obstacles.
This paper presents a low-level controller for an unmanned surface vehicle based on Adaptive Dynamic Programming (ADP) and deep reinforcement learning (DRL). The model-based algorithm Back-propagation Through Time and a simulation of the mathematical model of the vessel are implemented to train a deep neural network to drive the surge speed and yaw dynamics. The controller presents successful simulation results validating the feasibility of the proposed strategy and contributes to the diversity of validated applications of ADP and DRL control strategies.
The overall strategy, creative new elements, and final results from VantTec for RoboBoat 2020 are presented. The strategy relied on using experience from previous editions to continue working on top of the same platform, the VTec S-III USV, and have a higher focus on the software and scientific segment of the project. A simulation environment was developed to increase the testing time for the algorithms, and became the only source of testing as the COVID-19 pandemic grew. The new creative implementations include improved software architecture, user interface, and algorithms for path-following control, collision avoidance and visual-based guidance. Experimental results show the performance of the new path-following controller. Simulation results demonstrate the capabilities of the proposed solutions for each challenge based on local visual-based guidance and reactive collision avoidance. Finally, embedded systems and UAV mobile application sub-teams increased the performance of both technical aspects of the project, that had not been sufficiently worked on during previous RoboBoat editions.
Unmanned surface vehicles (USVs) must be able to safely navigate through unstructured environments. Perception and obstacle avoidance are essential capabilities for any USV related application. In many approaches for USV obstacle perception and avoidance, only simulation results are obtained. In this paper, a methodology for 3D vision based obstacle avoidance demonstrating experimental results is presented. This work integrates different 3D sensors, such as a stereo camera and 3D LiDAR. The methodology is validated with independent setups for its strategic components. Moreover, successful experiments with a physical platform are described.
In this manuscript, the design and implementation of a controller for a double thruster twin-hull unmanned surface vehicle prototype are presented. A three degree of freedom model is used to describe the ship dynamics, and given that the vehicle is underactuated, a controller based on the backstepping with constrains in the velocity is applied to drive the surge speed and heading variables. Then, to track desired trajectories, a guidance law based on GPS waypoints is adopted. Finally, experimental results on the platform demonstrate the advantages and properties of the proposed controller in two scenarios: surge speed and heading control and GPS navigation.
This technical design report addresses the strategy use by the group VantTec for the international competition RoboBoat 2019. This includes creating a modular software system to simplify challenge development, and managing members according to area of expertise and experience on the group. Furthermore , the creative decisions for new hardware and software subsystems, as well as experimental results on subsystems such as object detection, map creation and a backstepping controller.
This paper addresses the development of a three degree of freedom dynamic model for a custom-made unmanned surface vehicle, which includes an identification of parameters via measurements, estimation and through field experiments. Furthermore, a control approach based on backstepping with constrains to deal with the under-actuated nature of the model is designed in order to perform tracking of desired trajectories. Finally, the results validate the model by a comparison between experimental and simulation responses and shows the feasibility of the proposed control scheme.
Deep learning-based frameworks have been widely used in object recognition, perception and autonomous navigation tasks, showing outstanding feature extraction capabilities. Nevertheless, the effectiveness of such detectors usually depends on large amounts of training data. For specific object-recognition tasks, it is often difficult and time-consuming to gather enough valuable data. Data Augmentation has been broadly adopted to overcome these difficulties, as it allows to increase the training data and introduce variation in qualitative elements like color, illumination, distortion and orientation. In this paper, we leverage on the object detection framework YOLOv2 to evaluate the behavior of an obstacle detection system for an autonomous boat designed for the International RoboBoat Competition. We are focused on how the overall performance of a model changes with different augmentation techniques. Thus, we analyze the features that the network learns by using geometric and pixel-wise transformations to augment our data. Our instances of interest are buoys and sea markers, thus to generate training data comprising these classes, we simulated the aquatic surface of the boat and collected data from the COCO dataset. Finally, we discuss that significant generalization is achieved in the learning process of our experiments using different augmentation techniques.
This technical design report discusses the conceptual problems presented to the student group VantTec while preparing for the international competition RoboBoat 2018. First, managing the group members and the strategy to approach the competition with computer vision and GPS based navigation. Second, the design innovations on the mechanical and software subareas, and the overall system. Finally, the experimental results and the in water experiments to be performed before the competition.
Roboboat is an international competition where teams design autonomous boats to overcome challenges with the least possible error. The challenges evaluate the boat's ability to navigate autonomously, recognize objects, avoid obstacles and find paths to a destination. This paper documents the work done by the VANT Tec team during the development of the boat for roboboat 2017.