alex specific approach

Specific approach

For our experiment we will be having volunteers remotely operate (i.e, without vision of the robot) a PR2 or Turtlebot in a controlled environment. The user will be asked to perform some primary objective, such as "count the people in the library" or "find the books by author x". We will additionally implement "physical redaction" to protect various privacy concerns. An example would be telling the user that they are a fire-inspector robot who must verify the building follows a certain fire code. During this process we may want to hide stray cell phones and expensive computers that could be stolen, so we can block them from the robot's field of view with black paper.

In coordinating this experiment we will have to solve some technical issues, the first of which is network speed. Channeling a live video stream from one of our robots over a network is notoriously slow. To solve this we will tentatively create our own network in our environment for the purpose of improving the quality of the user's video stream.

After conducting the experiment we can compare whatever we have defined as our task performance metric categorically versus the control. If this metric is a Bernoulli variable (i.e, exclusively either "the user performed the task adequately" or "the user failed the task") then we can perform a difference of proportions t-test to determine if the groups are statistically different.

Specific techniques

• Physical Redaction -- Using cloth and paper to manipulate the environment physically. This can be verified by asking the user to violate the privacy concern.
• Create our own network to solve network lag issues.

Further research techniques

Once the above has been completed, there are additional improvements to explore, such as:

• Real time object removal. We can try to implement known filters to remove objects in real time from the video stream. This is also called "inpainting".
• Physical User Interfaces. Using color matching and then template matching to define physical markers that can be digitally recognized. This way you can, for example, use bright green paper to demarcate the boundaries of a door and then black out the contents therein.
• Improving localization. One yet untested technique is to hang cameras from the ceiling using easily recognizable markers to determine the position of the robot.
• Physical Privacy. Perhaps we can create a map of our environment and determine the area the robot preoccupies so as to prevent it from entering certain rooms. This would require minimal odometry accuracy.