ROBOT LOCALIZATION

SLAM (Simulataneous Localization and Mapping) is the core of all autonomous mobile navigation in the robotics space. A mobile robot operating autonomously in the environment runs four fundamental modules, namely perception, cognition, localization and mapping. Perception and cognition help the robot get information about the surroundings and use the same to generate control actions respectively. Localization and mapping refer to robot's placement in space with respect to a standard reference and the map of traversible and restricted paths in the environment respectively.

It needs synergistic operation of all these four modules to execute successful SLAM applications. Modern robotics techniques salvage the advances in machine learning and artificial intelligence in perception tasks like computer vision, odometry estimation, lidar data, sensor fusion and more. Cognition and robot control is largely conventional model-based techniques with some online model enhancement and techniques. AI-based control techniques have not transitioned from academia to the industry yet.

Mapping and localization are essentially two sides of the same coin. Mapping is largely done offline and then loop closure and feature matching is done during navigation for localization and mapping enhancement in the process.

The scope of this article is to discuss a general pipeline towards solving a localization problem.

INTRODUCTION

Of all the modules, localization is the most challenging due to high performance and high fidelity requirements for successful SLAM. Localization is largely dependent on sensors, mapping and the system response. For Mobile robots, localization refers to determining the pose of the robot, usually relative to the map but very often also relative to other components like humans and other static elements, in a dynamic environment.

Localization is a difficult task owing to several limitations in sensing technologies and environment dependencies. While localization in outdoor environments with GPS facility may be considered enough, it is often not for mobile robots. GPS is prone to poor resolution (like a few meters) and inaccuracies while autonomous robot navigation in unstructured environment needs very precise pose estimation. Outdoor applications like last-mile delivery, search-and-rescue, security need robot localization to be very accurate for successful task completion.

Additionally, mobile robots deployed in indoor applications, essentially GPS-denied environment like warehouses, retail stores and factories render the GPS obsolete. This leads to the requirement of a reliable and robust sensor suite. A typical mobile robot can have any or all of inertial, visual and distance sensors like lidar, ultrasonic rangefinder, camera, GPS, accelerometer, gyroscope, magnetometer and more.

As visible in the localization pipeline, the localization problem has two most important steps - the prediction step that guesses the pose of the robot based on the motion information (actuator encoder information used with the robot motion model) and the update step that utilizes the estimates of pose from external sensors and improves the result obtained during the prediction.

WHEELED MOBILE ROBOTS

As mentioned, despite of several limitations, the simplicity and large scope of applcations make wheeled robots the most favorable mobile robots. Wheeled robots are largely classified into two categories, holonomic and non-holonomic drives.

Holonomic robots are such that their constraints are all integratable to position constraints. It simpistic terms, the number of controllable degrees of freedom of the robot is equal to the total number of degrees of freedom. Mecanum and omni wheel drive robots are holonomic drives. They allow the robot to move in any direction instantly without any turning radius or realignment.

Non-holonomic systems are systems with controllable degrees of freedom less than than the total degrees of freedom. Ackerman, differential and belt drives are all such non-holonomic systems. Such mobile robots are incapable of instantaneous motion in a desired direction and always move on a circular trajectory.

Each of these drives have their own significance in indoor or outdoor applications. The forward and inverse kinematics models of these drives are used to control the motion. Forward kinematics establishes the relationship between the joints and physical space while inverse kinematics is needed to convert task space trajectory obtained from motion planning algorithms to joint space for motion control. For mobile robots, it is safe to restrict motion only in the XY plane with +X protruding out along the robot heading while the Y axis is perpendicular to the heading as per the right-hand rule.

Differential Drive Model

Differential drive is one of the most common drive mechanisms in delivery or warehouse robots. They have standard geometric wheels and one or more free passive wheels. The robot has a pair of parallel wheels placed on the same axis, on either sides of robot body. Relative motion in the wheels on either sides makes it move along different curves. Because of the relative motion, it doesn't need special powered wheels for steering. Instead, it just needs free wheels like castor or ball wheel that can align in any direction.

Because the differential drive is a non-holonomic drive, the differential drive kinematics defining its motion are not integrable to the final pose of the robot. The distances traveled by the two indvidual wheels are not enough and needs the time sequence of motion to obtain the final robot pose. The linear and angular velocities of the wheels are used to derive the robot position and orientation in the physical space.

Ackerman Drive Model

Ackerman drive is the special steering geometry that solves the problem of wheel slippage and difference in distance to be travelled by the inside and outside wheels while moving along a curve. All commercial vehicles are build on this drive mechanism. This mechanism has a powered pair of rear wheels while the front wheels are used to steering. The front wheels of the drive have a common instantaneous center of rotation that lies on the line joining the two rear wheels.

Mecannum Drive Model

Mecannum wheel drive is the special category of holonomic drive that can drive in any direction, instantaneously. The standard drive has 4 individually powered mecannum wheels. Mecannum wheels have small roller wheels aligned at 45° along the periphery an individual wheels. Different configurations of speeds and directions of the wheels make it possible to travel along specific directions.