Going through various developments, the mobile robots have come to the phase of Autonomous Mobile Robots. The general problem of mobile robot navigation was summarized into three questions: “Where am I?,” “Where am I going?,” and “How should I get there?”. In this project we tried to figure out these solutions to these questions by implemeting SLAM algorithms using ROS.
The question of ”Where am I?” addresses the localization problem of autonomous robot control. Localization involves the task of identifying the robot position with respect to its environment. In a typical robot localization scenario, the robot uses the sensor information and already available environment map. This approach takes in noisy sensor data that give the measured pose of the robot in the environment and applies various filtering approaches to find the estimate of the actual pose of the robot and solve the localization problem. Bayesian filtering can be employed as a powerful technique to address this problem. However, the constraint of already knowing the map weighs heavy on the this approach to the problem. The other approach overcomes this constraint in the sense it builds the map and localizes itself in it simultaneously and hence, called SLAM. The classical age of SLAM(1986-2004) witnessed the probabilistic approaches to SLAM among which were EKF and maximum likelihood estimation. The later period, algorithmic-analysis age (2004-2015), was much involved with the study of SLAM and its fundamental properties like observability, convergence and consistency itself.

SLAM as we have devised in this project requires laser-scan data coming from the laser range finder sensor or LIDAR attached to the mobile robot. The laser-scan data corresponds to the environment of the mobile robot and is used to develop the occupancy grid map of the environment and localize the mobile robot in it.
In practical applications, the laser data is received form LIDAR of the mobile robot and the SLAM processing is done resulting in the occupancy grid map. The obtained occupancy grid map gives the location of obstacles in the environment and that is used by path-planning algorithms to find the appropriate path from the robot’s current position to the specified goal position. However, as for our project we require the laser-data and generate the occupancy grid map. So we selected Robot Operating System(ROS) as our platform to simulate the mobile robot in an environment and gather the laser scan data and also to visualize the output i.e. occupancy grid map and the particle position in it. For the world simulation we use a 3D simulation package called Gazebo and readily available turtlebot 3 packages as our environment. The real time visualization of output is done using the package Rviz.

Particle Filter

Particle Filter uses a set of particles to represent the posterior distribution of a model in a process and filter out the particles as the number of observations increase, leaving out the particles with the highest probability of survival. In our project we used the particle filter for calculating the possible location of the robot in the given map, observing the laser scan data and filtering out the particles by calculating its importance weight.
At the beginning a list of particles is defined. Each particle has the necessary attributes like 2D position, orientation (θ) weight and their own individual map. The position of particles are then updated using the Hill Climbing algorithm, which uses random normal distribution and returns the best estimated position of the particles respective to scan data. After that the weight of each particle was calculated using the sensor model. The calculated weight of the particles are then normalized and re-sampled.

Tiny SLAM uses the hill climbing approach in order to search for the best possible position that matches the current laser scan data. We use a random hill climbing approach where, the algorithm iterates max number of times and returns the position for which the sensor model gives the least value. The robot poses are searched randomly by using a normal distribution. The parameters σxy, σθ can be used to control the variance of the search. Higher value of these parameters lead to a much larger search space.