Department of Computer Science and Engineering, C. V. Raman Global University
Highlights & Notes
Abstract
Localization: The location of a robot with respect to its environment.
Localization problem categories
Initial position of the robot
Position update principles
Key techniques to localize the mobile robot
Probabilistic approach
Markov localization
Kalman Filter
Autonomous map building
Simultaneous Localization and Mapping (SLAM)
Radio Frequency Identification (RFID)
Introduction
Performing successful navigation
Perception phase: extracting meaningful data by interpreting its sensors
Localization phase: robot estimates its current location in employed environment using information from external sensors.
Cognition phase: robot plans the necessary steps to reach the target.
Motion Control: lets the robot achieve it’s desired trajectory by modifying it’s motor outputs
Localization problem
Position Tracking
Robot’s initial position is known and the objective is to track the robot at each instance of time during its navigation in the environment
Localization algo utilizes the robot’s previous position to update its current location
Utilizes the odometry and sensor data
Position uncertainty of the robot is required to be small
Global position or localization
Robot does not have any knowledge about its initial position.
Robot is able to locate itself globally within the environment
Kidnapped Robot
Robot is abducted to an unidentified place
Robot knows it is being kidnapped
Kidnap recovery is necessary for any autonomous robots
Best match between known data and sensor data solve the issue of relocation.
It should be able to handle pose monitoring and relocation simultaneously
Static environment: Localisation is simple since robot is the only moving object
Dynamic environment: Localisation is significantly more difficult, because the robot gets confused about its location due to presence of other moving objects.
SLAM: In unfamiliar zones, the robot learns the environment’s map during navigation. Simultaneous Localization and mapping.
Localization Principle
A mobile robot keeps track of its motion using odometry
Odometry: estimating data from motion sensors to keep track of position over time
The external sensors along with robot’s odometry can be combined to localize the robot.
SBest – Robot’s belief
Robot’s position update has two steps
Prediction/action update
Here robot utilizes its proprioceptive sensors(e.g. Wheel/motor sensor, acceleration sensor) to estimate its position
Due to odometry error, the uncertainty increases and gets accumulated over time
Perception/measurement/correction update
Robot corrects the estimated position(outcome of prediction phase) using its onboard exteroceptive sensors
Initial position x0 is known, and hence the PDF is Dirac delta function.
Robot uses rangefinder to calculate its present distance d from the right wall and computes the position x’2
The position conflicts with the present position x2 estimated in the prediction phase.
Measurement update corrects the new location to x’’2
Consequently the uncertainty shrinks.
Updated position = Initial position + robot’s path in the last sampling interval
Localization Approaches
Probabilistic Approach
Identifies the probabilities of robot being in specific positions
Measurement errors affect the sensor data
Probability of a robot in a specific configuration can only be computed using
Markov Localization
Kalman Filter Localization
MARKOV Localization
Robot can locate itself starting from an unknown position
Multiple possible positions can be tracked by the robot
Can recover from ambiguous situations
State space needs to be represented in a discrete way(topological graph or geometric grid) for updating the probability of possible positions.
Memory requirement for map size is limited
Prediction update phase
Robot’s current location is estimated depending on the available info on previous locations and odometry input
KALMAN FILTER based localization
Solves the position tracking problem in efficient and precise manner
Optimal sensor fusion approach which tracks the robot from an initially known location
Special case of markov localization
Perception update phase
Step 1: Observation Step: robot extracts different features such as specific sensor value, doors, lines etc.
Step 2: Measurement/ Perception Prediction: consists of features that it expects to observe from its assumed positions that is from the outcomes of the prediction step.
Step 3: Matching: robot computes the best match between features extracted from observation and features selected during the measurement prediction.
Step 4: Estimation: updating the robot belief state
The robot initial position should be known with a certain approximation
If the robot gets lost, it can’t recover its position
It does not solve the global localization and the kidnapped robot problem.
Many robotic systems require the system to be nonlinear. This gives rise to another localisation approach commonly known as extended Kalman Filter (EKF), extension of Kalman filter to non linear systems
Here the system is linearised and then kalman filter is applied
Represents a Gaussian probability distribution
Simulated analysis show that, in comparison to KF, EKF performs better in mobile robot position estimation.
MCL: Monte Carlo Localisation
Uses a subset of the whole set of possible positions to construct the approximate belief state of the robot
Results in tracking and updating a small number of possible locations instead of all possible locations which reduces complexity
Object positioning system methods implemented by different authors
Vision based tracking system: Utilizes color stereo cameras
Active Badge System: Beacon based and RF schemes in office environment. Badge worn by every individual which emits a distinctive code in every 15s duration. Signal is received by a network sensor placed in a centralized location. Tag is responsible to determine the position of a person’s location.
Sound based localization: locating objects through sound
Complimentary filtering technique
Fusing data gathered from GPS, digital compass and IMU
Results in low implementation cost and high quality result for robot navigation in outdoor environment
TCP/IP Based client server system in intelligent space
Localizes and control s different types of robots including robot wheel chairs
System comprises of 11 cameras for synchronous image capturing
Server makes connection with each of these cameras and robot
Server collects various sensors data, processes them and sends the outcome to the client.
For Self localization, robot utilizes different landmarks such as beacons, walls etc.. Environment maps contain these landmarks
Strategy for autonomous map building
Robot starts navigating from a random initial point
Explores the environment autonomously through its sensors
Gain environment knowledge, builds a map through interpreting the scene
Helps the robot to localize relative to the map
Computer vision has helped a lot but creating and modifying the environment is still a major concern usually termed as SLAM problem.
Mapping creates the environment’s map from the information on robot’s accurate path.
SLAM utilizes the data collected from the robot’s proprioceptive and exteroceptive sensors to recover both the environment map and robot path
Collected data are based on the displacement of robot estimated from the odometry and features such as lines, corners and planes. Theses features are extracted from ultrasonic sensor, laser sensor and camera images.
SLAM is formulated in a probabilistic way, where the present position and robot map is estimated as a probability distribution.
The objective is to estimate the robot’s location and environment map from a given series of controls and sensor observation over distinct time steps.
Full SLAM: entire robot path is updated
Online SLAM: updates the robot’s current location only
EKF-SLAM based approach
Uses an extended state vector containing the robot pose and the position of all features in the map
Omnidirectional vision system
Vision system to produce environment information around the mobile robot. Info is used to extract the environment features.
Landmark is located and the EKF approach is used to synchronously update the attitude and location of the robot as well as map library.
SLAM based brain controlled robot localisation
Human EEG signal is recorded by 40 channels based digital EEG recording system (NuAmps-NeuroScan)
The motor imagery analyses this signal through the common spatial patter based support vector machine classification technique
The motor imagery (MI) analyzes this signal through the common spatial pattern (CSP)-based support vector machine (SVM) classification technique. The proposed SLAM scheme focuses on improved RGB-D SLAM which combines feature tracking (based on optical flow) and deep learning techniques. The proposed methods get speedup by optical flow strategy and object identification based on deep learning provides higher accuracy, stronger robustness, and additional environmental information Simultaneously. Human operator manipulates the robot through the BRI based on MI combined with SVM classification based on CSP. The suggested methodology locates the robot more correctly by removing the error of pose calculated from ordinary matching of features.
Control strategy with brain teleoperation
Technique adopts deep learning methods to navigate and control the mobile robot
Well suited in unknown environments
Robot is controlled by a steady state visually evoked potentials (SSVEP) based BCI system and deep learning.
Human intentions are recorded as EEG signals by NuAmps device
The online SVM classifered analyses EEG signals and decodes them to generate movement intentions which results in providing motion commands to control the robot.
Radio Frequency Identification Device (RFID)
Broadly classified into two categories
Active
Passive
Non battery requirement to maintain the wake and query cycle, the passive tags have lifetime limit
Active tags use battery
To localize a robot, set of tags are arranged in a grid pattern
Robot consists of RFID reader and an antenna(transponder)
Normally, RFID tags contains the location information. When the robot passes over a tag, the RFID reader attached to the robot detects the tag and extracts the position information
Hahnel Approach
Proposed a probabilistic measurement model for RFID readers that make it possible to locate RFID tags
System uses a laser range sensor to learn the environment’s geometric structure using FastSLAM algo
Position of the tags is estimated based on robot’s path
Here the authors applied MCL to determine the robot’s location globally
Localisation errow with and without odometry during global localisation with RFID tags is analysed.
Further laser based global localization error with and without RFID data is compared.
Analysis on Position and Orientation error
Position error is the difference between the robot’s real position and its calculated position
Orientation error is calculated as the modulus of the difference between actual and computed orientation.
Conclusions and future direction
Combining probabilistic localisation approach such as EKF and SLAM increases localisation and orientation accuracy, which in turn reduces the positioning error.
Further in low noise environment, EKF method is well suited to provide precise solution with higher convergence rate
In limited environment space such as indoor environments, the RFID scheme is proved to be very efficient in tracking robots
Apart from SLAM efficient continuous localisation and component mapping schemes like ARIEL and feature based extraction may also be combined with mobile robot localisation techniques.
