Need for Robotics in Manual Scavenging and a Proposed Methodology

2020, Oct 23    

A deadly practice of Manual scavenging is common in South Asian countries. Manual Scavenging refers to manually cleaning, carrying, disposing human waste from septic tanks, sewers and dry latrines. In India alone 1.8 million people are forced to engage in these activities because of socio-economic reasons. Thousands of scavengers are dying around the world while doing so, out of which manhole scavenging is worse. My research topic focuses on creating a low cost efficient robot manipulator and its peripherals so that we can replace humans. Further the research methodologies gained from this project can be used in other projects of Ocean water cleaning, underwater navigation and manoeuvring robots in difficult terrain.

Keywords: Robotics manipulators, Reinforcement learning, Deep learning, SLAM

Research objectives

The existing Manhole cleaning process requires to manually position the sewer rods, sewer jetting machine into the manhole to clean sewer lines[1]. But these machines are not effective in cleaning the lower part of the manholes which have Hard rocks, Sand, Silt, Sludge, Non-Biodegradable wastes, etc. Because of the ineffectiveness of these methods a worker needs to enter a manhole, which is filled with toxic gases and infected liquid waste to open blockage. Things became worse during the time of COVID-19[2].

The research project can be broken down into these main tasks:

• Robotic manipulator with enough degree of freedom and length to pick up the waste from the bottom of manholes.
• Water resistant and dust proof body. Also the body should be made up of lighter, stronger material which is corrosion free.
• Computer vision using a camera, which can operate in dark conditions.
• AI techniques to collect and process data and can navigate inside manholes.
• Detection and alert for harmful and toxic gases released during fragmentation of waste.
• A GUI in multiple languages so people of different countries can operate this robot.

Methodology

Note: Methodology I am proposing is based on my knowledge of robot designing and literature review. I thoroughly believe that modification in this approach is liable.

Designing a manipulator arm is the heart of our problem. There are several constraints attached with this. Manholes are typically 30-45 inches in diameter, our robotic arm should be designed to enter the small hole by converging and later expanding accordingly. Manholes can be upto 15 meter in depth, so our arm should be adjustable length to serve the purpose.

System architecture for this is going to be a controller and Robot. At first the robot can be a simulator and later hardware. For the controller portion we need to design a supervisory logic to calculate where we want our end effector to be. In the manipulator algorithm there is a notion of forward and inverse kinematics. Inverse kinematics is calculating a given position and orientation relative to the start of the chain. Inverse kinematics is more difficult to calculate. One of the methods to solve it is using Denavit–Hartenberg parameters approach. But this is computationally expensive compared to new machine learning optimization techniques. This paper [3], tells a deep reinforcement learning approach for dynamically stable inverse kinematics of humanoid robots. Their proposed strategy was evaluated on the upper body of a humanoid having 27 degrees of freedom. Their accuracy was about 90% after 9900 episodes. However I believe that these episodes can be reduced to lesser numbers and more accuracy can be obtained by following this paper[4], by Google Brain Team. They proposed the idea of moving robot reinforcement learning to the simulation first, because it will be more scalable and easy to operate. This was good for getting a lot of data but the model trained in the simulator tended not transfer to actual real world robots. There are a lot of systematic differences between two, one is visual appearance, other is real world physics and stimulated physics. They proposed the way to improve this by adding more randomization into the simulator. They particularly used Domain Adversarial Neural Network (DANN) to do so. Hence by combining the results from paper[3][4], we can achieve greater accuracy in less number of episodes.
The other aspect of the problem was localization and mapping. In computer geometry, simultaneously location and mapping(SLAM) is the method for construction of maps of unknown environments while also localizing itself in the environment. In a manhole, whose inside depth and width is unknown this method is appropriate to approximate the map. Below are some results of SLAM of Manhole, Figure 2 & 3. Decomposition of organic matters in sewers, manholes and pits of the drainage system will generate methane and/or hydrogen sulphide. Hydrogen sulphide, being very soluble in water, often dissolves in sewage and can be trapped within sediment and sludge in sewers as gas pockets. Entering these manholes is dangerous. That's why mapping from robot device is important[7].