In order to collect good statistics, I will need to design experiments to test my software. I think this will consist of a process of running AMCL on a robot and driving it through a known environment.
Possible Experiment Design
While AMCL is running, my CNN Localizer will run and continually run images to give an estimate of the current position / region of the robot.
While monitoring the covariance of the AMCL pose cloud, a program will run to check the AMCL pose versus the region identified by the CNN Localizer. The CNN Localizer's estimate of position will be checked against AMCL's and if they are in agreement, that 'trial' will be marked as a success. Otherwise, it will be marked as a failure and added to a failure report.
The report will contain all of the instances in which a failure occurred along with AMCL's estimate pose, the covariance of the particle cloud, as well as the image run through the network with its true label and output label.
Experiment Design
In order to collect good statistics, I will need to design experiments to test my software. I think this will consist of a process of running AMCL on a robot and driving it through a known environment.
Possible Experiment Design
While AMCL is running, my CNN Localizer will run and continually run images to give an estimate of the current position / region of the robot.
While monitoring the covariance of the AMCL pose cloud, a program will run to check the AMCL pose versus the region identified by the CNN Localizer. The CNN Localizer's estimate of position will be checked against AMCL's and if they are in agreement, that 'trial' will be marked as a success. Otherwise, it will be marked as a failure and added to a failure report.
The report will contain all of the instances in which a failure occurred along with AMCL's estimate pose, the covariance of the particle cloud, as well as the image run through the network with its true label and output label.