EKF in MS Robotics Studio 2010 - Revision history

Robot Lab 2010 at 00:00, 20 March 2010

2010-03-20T00:00:22Z

Robot Lab 2010: /* Extended Kalman Filter */

2010-03-19T23:59:21Z

Extended Kalman Filter

Robot Lab 2010: /* Conclusion */

2010-03-19T23:57:38Z

Conclusion

Robot Lab 2010: /* Conclusion */

2010-03-19T23:57:13Z

Conclusion

Robot Lab 2010: /* Conclusion */

2010-03-19T23:55:36Z

Conclusion

Robot Lab 2010: /* Data */

2010-03-19T23:53:00Z

Data

Robot Lab 2010: /* Data */

2010-03-19T23:52:51Z

Data

Robot Lab 2010: /* Data */

2010-03-19T23:52:22Z

Data

Robot Lab 2010: /* Media */

2010-03-19T23:51:50Z

Media

Robot Lab 2010: /* Analysis */

2010-03-19T23:50:39Z

Analysis

@@ Line 23: / Line 23: @@
 == Assumptions ==
 In order to make the simulation easier and get to the core issues of implementing an EKF-SLAM algorithm, I made several assumptions. They are listed below.
-* The random moving my robot does has two seperate phases. Move Forward and Turn. Every second, I either move forward for a random distance or turn a random number of degrees. During the forward phase (this is always the first phase to go), I move forward, update the location based on odometry, and perform the Time Update Step of the EKF. During the degrees phase, I rotate, update the direction the robot is facing, and perform the Observation Update Step by observing the landmarks in the environment. In practice, the robot would most likely continuously move, but for the sake of keeping the simulation easy, I chose the move in steps.
+* The random moving my robot does has two separate phases. Move Forward and Turn. Every second, I either move forward for a random distance or turn a random number of degrees. During the forward phase (this is always the first phase to go), I move forward, update the location based on odometry, and perform the Time Update Step of the EKF. During the degrees phase, I rotate, update the direction the robot is facing, and perform the Observation Update Step by observing the landmarks in the environment. In practice, the robot would most likely continuously move, but for the sake of keeping the simulation easy, I chose the move in steps.
 * Instead of adding landmarks each iteration, I kept the number of landmarks static throughout the run.
 * I also did not do any data association. I felt this would really complicate the problem and push me away from my focus of the Kalman Filter.
@@ Line 32: / Line 32: @@
 ** SimulatedSLAM
 *** This is the main class for the Service and is essentially the driver for my program
-*** Performs Odometry and uses Normal Distribution to create noice in the odometry data
+*** Performs Odometry and uses Normal Distribution to create noise in the odometry data
-*** Performs Landmark indentifying and uses Normal Distribution to create noise in the data
+*** Performs Landmark identifying and uses Normal Distribution to create noise in the data
 ** SimulatedSLAMState
 ** SimulatedSLAMTypes
@@ Line 68: / Line 68: @@
 # True Location - This is the exact location of the robot. The odometry is perfect. I mainly record this since the MRDS Simulation Engine returns all data without noise and then I add the noise myself.
 # Odometry-Based Location - This is the location of the robot based only on odometry with noise.
-# SLAM-Based Locaiton - This is the location of the robot based on odometry with noise and landmark identification with noise.
+# SLAM-Based Location - This is the location of the robot based on odometry with noise and landmark identification with noise.
 === Media ===
 To view media (videos and screen shots) [http://rise.cs.drexel.edu/~mcd45/kalmanfilter/media/ see here].
@@ Line 146: / Line 146: @@
 General knowledge computer science terms are not defined here.
 * MRDS - Microsoft Robotics Developers Studio
-* CCR - Concurrency and Coordinatio Runtime
+* CCR - Concurrency and Coordination Runtime
 * DSS - Decentralized Software Services
 * SLAM - Simultaneous Localization and Mapping
@@ Line 153: / Line 153: @@
 == Hints for Starting and Using MRDS ==
 One of the main hurdles I encountered when doing this project, was the learning curve of MRDS. Here are some hints for anyone interested in programming in the MRDS environment.
-* Just because you are familiar with the .NET environment, does not mean MRDS will come easy to you. I had been programming .NET Web Application for the past 6 months when I started the project and I thought it would be very similiar. While a lot of the code and basic libraries that are the same, there are many practices, such as concurrency, that is not used in ASP.NET. With this said, give yourself plenty of time to learn the basics of the MRDS architecture (CCR, DSS, Servies, Simulation Environment, etc.)
+* Just because you are familiar with the .NET environment, does not mean MRDS will come easy to you. I had been programming .NET Web Application for the past 6 months when I started the project and I thought it would be very similar. While a lot of the code and basic libraries that are the same, there are many practices, such as concurrency, that is not used in ASP.NET. With this said, give yourself plenty of time to learn the basics of the MRDS architecture (CCR, DSS, Services, Simulation Environment, etc.)
 * I did not find many great tutorials out there for MRDS. Most of the tutorials on the [http://msdn.microsoft.com/en-us/library/default.aspx Microsoft MSDN Library] where not helpful when it came to driving a robot through a simulated environment. The tutorials covered Services in a general sense and covered very basic Simulation practices. I was not able to find a tutorial on there site that gave code for driving a vehichle. I luckily found an example at [http://promrds.com/ ProMRDS] that gave some very helpful code. You can also look at my [[EKF_in_MS_Robotics_Studio_2010#Code|code]], which has a fairly straightforward implementation of randomly moving the vehicle based on time ticks and keeping track of the location based on odometry.

@@ Line 8: / Line 8: @@
 === Extended Kalman Filter ===
 An EKF (Extended Kalman Filter) is a Kalman Filter that can run with non-linear equations in addition to linear equations. This is where the "Extended" comes in to the name. SLAM is considered a non-linear problem, and therefore an EKF is needed. The EKF does not actually deal with the non-linear equations, it instead "linearizes" them through the use of Jacobians.<br/>
-Like a regeular Kalman Filter, EKF has two main steps: Predict and Update. During the Predict stage, the next state vector <tt>x</tt> is predicted based on current state vector, control input and <tt>F</tt> jacobian. The estimate copvariance is also predicted. During the Update stage, the Kalman gain is calculated, which is used to update the state vector predicted in the previous step and the covariance is updated base on the <tt>H</tt> jacobian. The <tt>F</tt> jacobian is the partial derivatives of the function used to calculate the next state from the current state. The <tt>H</tt> jacobian is the partial derivatives of the function used to predict the measurement based on the predicted state [http://en.wikipedia.org/wiki/Extended_Kalman_filter]. For specific equations [http://en.wikipedia.org/wiki/Extended_Kalman_filter#Predict_and_update_equations see here].
+Like a regular Kalman Filter, EKF has two main steps: Predict and Update. During the Predict stage, the next state vector <tt>x</tt> is predicted based on current state vector, control input and <tt>F</tt> jacobian. The estimate covariance is also predicted. During the Update stage, the Kalman gain is calculated, which is used to update the state vector predicted in the previous step and the covariance is updated base on the <tt>H</tt> jacobian. The <tt>F</tt> jacobian is the partial derivatives of the function used to calculate the next state from the current state. The <tt>H</tt> jacobian is the partial derivatives of the function used to predict the measurement based on the predicted state [http://en.wikipedia.org/wiki/Extended_Kalman_filter]. For specific equations [http://en.wikipedia.org/wiki/Extended_Kalman_filter#Predict_and_update_equations see here].
 === EKF-SLAM ===

@@ Line 132: / Line 132: @@
 == Conclusion ==
-An EKF is a good way to solve Localization and Mapping. The algorithm I implemented is a very basic EKF-SLAM without the more advance techniques (data association, etc.) With that said, I was still able to show an improvement over the odometry. Future parts to implement would be data association and a more advance simulation environment, as this one was fairly simple. During this project, I learned a lot about localization and mapping, SLAM, and Kalman filters, in addition to improving my .NET skills greatly. This page should be a good start to anyone looking to implement a Kalman filter or work in the MRDS environment.
+An EKF is a good way to solve Localization and Mapping. The algorithm I implemented is a very basic EKF-SLAM without the more advance techniques (data association, etc.) With that said, I was still able to show an improvement over the odometry. Future parts to implement would be data association and a more advance simulation environment, as this one was fairly simple. During this project, I learned a lot about localization and mapping, SLAM, and Kalman filters, in addition to improving my .NET skills greatly. This article should be a good start to anyone looking to implement a Kalman filter or work in the MRDS environment.
 == References ==

@@ Line 132: / Line 132: @@
 == Conclusion ==
-An EKF is a good way to solve Localization and Mapping. The algorithm I implemented is a very basic EKF-SLAM without the more advance techniques (data association, etc.) With that said, I was still able to show an improvement over the odometry. Future parts to implement would be data association and a more advance simulation environment, as this one was fairly simple.
+An EKF is a good way to solve Localization and Mapping. The algorithm I implemented is a very basic EKF-SLAM without the more advance techniques (data association, etc.) With that said, I was still able to show an improvement over the odometry. Future parts to implement would be data association and a more advance simulation environment, as this one was fairly simple. During this project, I learned a lot about localization and mapping, SLAM, and Kalman filters, in addition to improving my .NET skills greatly. This page should be a good start to anyone looking to implement a Kalman filter or work in the MRDS environment.
 == References ==

@@ Line 132: / Line 132: @@
 == Conclusion ==
 == References ==
 The references below are to papers. All web references are cited and have a link where the citation is.

@@ Line 126: / Line 126: @@
 |}
-To view data and graphs, [http://rise.cs.drexel.edu/~mcd45/kalmanfilter/data/ see here]
+To view data and graphs, [http://rise.cs.drexel.edu/~mcd45/kalmanfilter/data/ see here].
 === Analysis ===

@@ Line 125: / Line 125: @@
 |7.889,1.217,5.750
 |}
 === Analysis ===

← Older revision		Revision as of 23:50, 19 March 2010
Line 129:		Line 129:

	=== Analysis ===		=== Analysis ===
		+	By looking at the data from the chart, you can see that the SLAM-Based approach produces a better location estimate than just the odometry. This shows that by including landmarks in the algorithm, improves the accuracy of the localization. However, it does not seem that SLAM algorithm is converging. Much like the odometry, the error seems to get larger and larger. The error is still smaller than just odometry though. There are several explanations as to what is happening here. It could be that the algorithm has not been run long enough and if it is run longer then the divergence will be much clearer. Another reason is that there is something not working correctly. The latter is tough to figure out, because Kalman filters are notoriously tough to debug.

	== Conclusion ==		== Conclusion ==