Robust Sensor Selection for Sidewalk Obstacle Avoidance using Reinforcement Learning

Sensors for assistive systems development to avoid sidewalk obstacles require certain qualities such as the sensor data acquisition must be ambient light independent so that it works both in day and night. In addition, the sensor's field of view (FOV) must cover the area of a typical sidewalk. Furthermore, the sensors should be energy efficient and smaller in size. The only way to find out the best sensor satisfying these qualities is to plug in the sensors to prototype devices and perform experiments. This process demands time and resources. In this research, we examined a robust sensor selection process by robot operating system (ROS) simulation. This simulation involves the reinforcement learning (RL) algorithm. We created a virtual sidewalk in Gazebo and let a robot with a sensor learn to avoid obstacles through RL. We found that Intel RealSense is best among the sensors which has optimum FOV and depth. We also designed a virtual sensor with time of flight (ToF) technology to study the light dependency, cost, and energy efficiency.