IoT Enabled Indoor Autonomous Mobile Robot using CNN and Q-Learning

Abstract

This paper focuses on the construction of IoT enabled mobile robot with an arm which can reach the destination autonomously and perform suitable actions in an indoor environment. Object detection and optimal navigation are the required features of a mobile robot that will be achieved through the combined architecture of building necessary Deep Learning and Reinforcement Learning models workable in a lesser memory space. To facilitate navigation in an indoor environment, initially, the environment has been mapped into a minimum number of grids for the experimental purpose. For handling huge memory requirement to run the models for processing, we occasionally transfer required intelligence from cloud setup to RPi, where RPi act as a Fog node in Industry 4.0 environment. The practicality of the robot has been gauged in three different cases (i) where the destination of the robot is known with 100% probability, (ii) where the destination of the robot is uncertain i.e. with lower probability and (iii) the destination is not known. In the first two cases, the objects are assumed to be stationary. Whereas in the third case, the objects can also be dynamic i.e. moving objects. As an application we have chosen Indoor Plant Monitoring System, where the objective is to measure the readings like Soil Moisture, Temperature, etc., of the indoor plant and forward the readings to Ericsson's IOT Accelerator platform. After analyzing the sensor values, a robot arm can initiate specific actions on its own. Here, the application of AI algorithms will not only help the robot to reach the destination, but it also triggers the robot to perform the functions optimally. As an experiment, we have studied the effect of learning rate on the total number of actions and introduces optimal reward from start to end of a journey in $4\text{X}4$ grid world environment and finally tested for tangible performance towards navigation and object detection.