PRE2022 3 Group9

Eryk Oyku Siiri Matilda Nika Maud

Topic: We want to do something more prototype based

Everyone must come up with sources of research for the Deep Sea Passive collection robot

Problem statement and objectives: We want to design an autonomous robot for deep sea garbage collection.

Users: Government organizations that focus on the environment.

What do they require:

Approach, milestones and deliverables: start with literature research to see what already exists, identify points that research is still needed and propose a way to combine multiple (existing) technologies to create a working robot.

Who's doing what?:

Matilda: image recognition model with TensorFlow

- Sort training data by size/looks of plastic

- Goal: 100 pictures

- https://www.tensorflow.org/tutorials/images/classification

Maud: I found the following sources, which, based on their abstract, are about under water waste collection or surface waste collection. When I have read them I will add a summary.

• An FM*-Based Comprehensive Path Planning System for Robotic Floating Garbage Cleaning. DOI: 10.1109/TITS.2022.3190278
  • This article discusses a method for surface garbage cleaning robots to find a good route to collect all the garbage in an environment that contains obstacles. It is assumed that it is known where all the garbage and obstacles are located.
  • For the path planning, first the order is determined in which the robot will visit all the pieces of garbage and then the route is determined.
    • To determine the order in which the robot will visit the garbage, the problem is modeled as a Traveling Salesman Problem. However, instead of the Euclidian distance, that does not take the obstacles into account, the authors used a heuristically guided FM* based distance (FM* stands for fast marching, I don't know exactly what FM* distance is, but it makes sure that the path goes around the obstacles, creates a smooth path and is fast to compute. FM* is similar to Dijkstra, so it computes the shortest distance form one point to the other, but it does not uses a partial differential equation to estimate the distance).
    • For the route of the robot, it is not always possible to chose the shortest path, because the robot needs to keep enough distance to the obstacles to not accidentally bump into them if they move and the robot might not be able to follow all paths, for example because it is not able to make a curve smaller than a certain diameter (the angle is to small). To make sure that the robot does not need to make to steep curves, a Gaussian filter was applied. The bigger the sigma in this filter, the bigger the curves and the further the robot stayed away from the obstacles. (There are formula's in the article about exactly how they applied the Gaussian filter, but I could not follow exactly what they were doing).
  • During the execution of the plan, the robot can get new information about the location and movement vector of the garbage it is currently heading for. If the garbage has shifted due to the current, the robot uses a trained neural network to determine its new route. For as far as I understand, the location of the obstacles and possible movement of the obstacles is not taken into account. The system does not re-compute the order in which it collects the garbage, even if the new situation as a different optimal collection order.
  • There are a lot of formulas describing what they do, but they often do not make it clearer. If we need them we can look at them later.
• A modified YOLOv4 detection method for a vision-based underwater garbage cleaning robot. https://doi.org/10.1631/FITEE.2100473
  • This article discusses an algorithm for detecting and classifying garbage underwater and determining its location compared to the robot.
  • This detection needs to happen fast and in real time, because the robots environment will be constantly changing.
  • The researchers use the YOLOv4 network, which is a one stage neural network that takes in the image and returns both the class and location of the objects in the image. The researchers pruned to reduce the number of calculations needed.
  • The network was trained on a data set, where it had to distinguish nets, plastic bags and stones. The data consisted of images from multiple sides and in multiple conditions made in a swimming pool. For the robot to perform in the real world, it needs to be able to detect more different kinds of garbage and also things like fish and under water plants. More realistic images could also pose a problem for this data set.
  • Pruning is used to significantly increase the detection speed. Pruning is achieved by not calculating the channels with the smallest contributions. This method works now, because there are very few categories, however, for a more realistic model, this would not work anymore.
• Open-Frame Underwater Robot Based on Vector Propeller Control. DOI: 10.1109/NetCIT54147.2021.00044
• DAMONA: A Multi-robot System for Collection of Waste in Ocean and Sea. https://doi.org/10.1007/978-981-16-8721-1_15/
• Design of water surface collection robot based on deep sea cage culture. DOI: 10.1088/1742-6596/2229/1/012005

Nika:

• Review of Underwater Ship Hull Cleaning Technologies | SpringerLink (tue.nl)
  • This paper presents a comprehensive review and analysis of ship hull cleaning technologies. Various cleaning methods and devices applied to dry-dock cleaning and underwater cleaning are introduced in detail,
  • Using the analysis of these technologies, we could take the positives and negatives into account in our research.
• Analysis of a novel autonomous underwater robot for biofouling prevention and inspection in fish farms | IEEE Conference Publication | IEEE Xplore (tue.nl)
  • Biofouling is a challenge for finfish farming as it can impact cage stability and fish health. Amongst others, current strategies against biofouling rely heavily on removal of biofouling using in-situ pressure cleaning of nets. The cleaning waste is released into the water where it can impact the health of the cultured fish
  • We can take the health of fish into account when designing the robot

Siiri:

• Plastic Waste is Exponentially Filling our Oceans, but where are the Robots? | IEEE Conference Publication | IEEE Xplore
• Robotic Detection of Marine Litter Using Deep Visual Detection Models | IEEE Conference Publication | IEEE Xplore
• Deep learning-based waste detection in natural and urban environments - ScienceDirect