Project 16

Grounding via Embedded Cues and Affordances Utlizing an Android Phone as a Low-Cost Robotics Platform.

Should this project be considered for Best Project Award: no

Should this project be considered for top 3 project awards: no

Project Organization/Clarity: 8/10

Project idea: 9/10

Project Contribution: 7/10

The project report was written quite clearly but there were a couple things that were seemingly unclear.  The use of an Android phone was a reasonable way to make this system optimally robust and low cost.  I was left somewhat unclear about how the simulation would take place.  More specifically, the navigation part of the system was, at times, listed as unsuccessful but i didn’t perfectly understand how that navigation was meant to take place.  I’m imagining some kind of routing around the home but I saw no real text to support this vision.  Notable mention was made into considered systems but without an explicit goal stated.
    On the same note, I didn’t understand much of the physical execution and felt like that was somewhat understated.  The operation section is marked as a sequence of modes that are, as I understand it, executed in succession to make up this system.  I’m somewhat unclear as to whether this system was executed sequentially or different parts were active at the same time (learn a little, go into other modes, come back and learn some more, etc).  I also wonder if attempts were made to relearn certain tags if that was one of the failure poiints or if the commands were recreated once a command did not perform adequately.  Mention was made that learning rate wasn’t a goal metric of the system but would be relevant in real world applications.
    The part i liked best about the project wast the idea in general.  Obviously, there has been extensive research into object recognition with the eventual goal of being able to create a chore robot.  This research proposes a system that largely bypasses difficult recognition tasks in favor of creating a function system.  These types of controlled environments are becoming more and more viable with new technology and such playground systems are at the point where its almost viable to implement the system in the home.  This is somewhat similar of the setup they have at the smart-home lab in the Computer Science department here at Iowa state.  I think its impressive that one person made a system that is analogous (albeit simplified) to a system that is an active research of technology for a whole department.
    I was somewhat put off by the number of technological workarounds that had to be implemented and nested in the final product of the system.  In its core, this system proposes a workaround to robust object recognition and exploration by embedding tags in the system.  Ideally, a robot would be able to identify objects on its own and use exploration to find the affordances of the objects.  Defining the affordances beforehand was a good approach to avoiding systems that are basically infeasible in the current time.  On top of that, however, tags were now switched out for QR codes making the system additionally nonrealistic.  Granted that doesn’t really matter since the system was implemented without a real robot so actually embedding something like RFID into objects would be overkill for this project in this scope.  Beyond that, however, it just felt like a bunch of systems that were part of the original plan for this project were passed up for inferior system for a variety of reasons.
    With the exception of a few changed out systems, this result was pretty consistent with the ideas proposed in the project proposal.  The idea was, furthermore, largely successful with the simulation giving proper responses to vocal commands most of the time.  Looking through the commands, however, a number of questions pop up.  You seem to have encoded a number of objects of the same type (i.e. 5 plates).  are all chair tags unique or could you essentially learn the same plate a bunch of times and it would count it as a new plate each time?
    I don’t believe there are major implementation details left out of the report but there are seemingly some small things missing.  The encoded record mentioned in the object table don’t seem to be perfectly standardized, was parsing handled individually to account for different changes or was some data just not deemed important making it only necessary to parse the parts that were standardized?  As I mentioned before, i would have liked to know the representation of the navigation used.
    Overall, the results reported do seem successful as most command responses produced the desired output.  A couple of the commands did not produce all the output desired but a perfect system was unfeasible to implement given the scope of the project.
    Future work and the work it takes to make this more publisheable are pretty well expressed in the paper itself.  It would probably be necessary to remove the components that made the system low cost.  That is, the QR codes would probably have to be changed out for RFID tags are other tags that are not as intrusive.  Ideally, the encoding could be done based on simple object title and object recognition based on vision.  That is, the system would get an image of an object and would be able to identify it, store its location and extract functional components.  A very far-off system would be able to learn all the affordances on the fly or during some babbling phase where the robot would have free reign to explore the house.  For the current system, however, the use of a better APIs would probably suffice in makign this work publishable.