Project 11. Learning Manipulation of a Flashlight.

Should this project be considered for the Best Project Award? Yes.
Should this project be considered for the top 3 project awards? Sure.

On a scale of 1 to 10, how would you rate the overall organization/clarity?
9/10

On a scale of 1 to 10, how would you rate the overall project idea? 
9/10

On a scale of 1 to 10, how would you rate the overall research
contribution of the project idea, methodology and/or results?
8/10

The final report is a good read. It was very nice to see a clear description of the performed work and renewed focus on related work. My only remark is that some of the images in the report could be a bit bigger.

It was also great to see a huge leap in quality between the proposal and the final report.

For future work, the idea with illuminating movement can be feasible with application of machine learning algorithms. For instance, the same buttons paper [2], which is cited in the report, contains a description of the active learning technique used to acquire the button pressing skill in real time. The idea was to use k-NN active learning in order to learn to move the arm in a pre-specified task-space pyramidal region. A similar approach can be used with the flashlight -- given the task-space region the robot trains a model in real time that allows it to select a candidate behavior out of a number of random candidates so that the flashlight illuminates what it has to illuminate.

Some background with Particle Filters and general familiarity with Sequential Monte Carlo and Markov Chain Monte Carlo techniques such as the Metropolis Algorithm can be useful for programming these tasks.

On the real-time side, the robot control program can communicate video data to a plugin, where a more high level logic can be placed. Thus, learning can be performed on a robot in real time. It is also possible to use Weka machine learning library through a JVM layer or using a sub-process.

One more thing I'd like to add to the related work discussion is a somewhat more philosophical flavor. What is a flashlight? How can a robot understand what it is doing with a flashlight?
What does it mean to use a flashlight? What does it mean to use a flashlight for a robot? What does it mean to mean in general?
How can the robot really know that the flashlight enhances its visual capabilities? How can the robot acquire a concept of a flashlight or, perhaps, a microscope? Can the robot generalize its knowledge of a flashlight to, say, a Geiger counter? What is really common between these tools? How can this knowledge be acquired from experience? Can language be used to communicate the idea of a flashlight to the robot?

Some of these questions have to be posed and perhaps tackled before autonomous intelligent robots start performing search and rescue missions with or without flashlights.

Keep up the good work!