For my “Computing Practice and Theory” project, I chose to continue some of the work with Prolog and EBNF grammar parsing that I had begun in the “Computability” program.  Specifically, I wanted to study more about Natural Language Understanding– the process of gleaning intended meaning from natural language.

As Natural Language Processing can be a bit dry, I decided to put a fun spin on the project; I designed and implemented a small command grammar for a Scribbler II robot.  The Scribbler II is a simple educational robot which has a python-based interface library called Myro available.

My program, tentatively titled “ScribPro”, is a hybrid python/prolog application that allows control of a Scribbler via natural language.  All communication with the robot is transmitted over Bluetooth.  The python component handles initialization, messaging to & from Prolog, code execution, and both fetching and sanitization of user input.  I was learning python on the fly as I did this project, so the final python component is pretty messy… but by the end of the quarter I felt that I have achieved basic competency.  I am very glad that I learned it; the ability to use python quickly create simple programs or prototype has already proven itself invaluable.

The Prolog component handles the bulk of the parsing work.  It reads the sentence fed to it from the python component, determines whether or not it is a successful parse, and if it is, converts the parse tree into a functor which it passes back to python.  The validity of a parse is determined according to a grammar.  Early on in the project, I tried to write a very abstract grammar using only the rules of the English language, and quickly realized that it was beyond the scope of this project.  This dilemma is one that anyone trying to do natural language parsing inevitably encounters: English is a messy language!  In order to overcome this and make a usable program within a ten-week timeframe, I chose to severely bound the words and sentential forms that my program would recognize.  Limiting the scope of my grammar to natural language units that are applicable to robot control made the project feasible.

While the final grammar never grew quite as complex as I would’ve liked, I am pretty happy with how the project turned out.  The robot is able to parse & execute a variety of movement-based commands — from plain English!

Not only is the robot able to handle simple sentences such as “Please go forward”, it can handle fairly complex compound sentences such as:

• “Spin in a circle to the left two times.”
• “Go forward four point five feet and turn left, then wait nine hundred milliseconds, turn around and beep once.”
• “Scoot forward until you reach the wall, then move backwards two inches, turn around and beep.”

Beyond the basic text mode, there is an optional voice input mode that taps into the Microsoft Speech Engine.  It is pretty fun to vocally order the robot around!

The fact that I was using two programming languages in tandem (prolog & python) presented some significant initial difficulties in the project.  A good portion of my final code revolves around passing messages back and forth between the two.  The tradeoff was worth it, however– had I been forced to implement my program in a single language, it would have been quite a bit messier.  There is something to be said for using the right tool for the job.  By the end of the project, I gained a substantial amount of confidence in my own ability to multiple languages within a single project.  No doubt this will come in handy in the future.

Here is a diagram of the program flow and component design of my final program:

I now feel confident that I have a handle on the basics of natural language parsing using grammars, and believe that I could implement such a system again in a much shorter timeframe, with a more elegant end result.

If I were to revise this project, I would attempt to abstract the code-generation side to be a level higher.  Right now, the code generated from a successful parse is extremely Scribbler-specific… which I would like to rectify.  I could conceivably create my own generic set of common robotics-related functions which the parser would invoke, rather than having it call specific Myro functions.  I would then implement the ability to map these generic functions to model-specific ones via config files; essentially allowing the parser to be used for other types of robots with ease.

Overall, I had quite a bit of fun with this project, whilst still managing to feel like I challenged myself.  I look forward to further independent research into Prolog, as I feel that logic programming is a valuable yet underutilized programming paradigm.  I also look forward to further tinkering with more advanced robotics projects.

Past weekly blog updates can be found at the following links:

• Week 1: Project Proposal – First Draft
• Week 2: Revised Project Proposal
• Week 3: Final Project Proposal
• Week 4: My Scribbler II Has Arrived!
• Week 4: A Sample Vocab List
• Week 5: Adventures In Integration
• Week 6: Finishing the Python/Prolog Interface
• Week 7: Scribby’s First Steps
• Week 8: You Can Walk The Walk, Now Talk The Talk
• Week 9: Cleaning House (And Code) — includes the full grammar definition.

My final code can be found here:

• scribpro.py – Python Main Program Executable
• nlp.pl – Prolog Natural Language Parsing Component

(Please be forgiving of the code, as I am an undergrad and was learning by doing for most of it.  There are approximately a million things I would change if I could do it all over again.)

If anyone reading this has any questions or is working on a related project, I’d love to hear from you!

I accomplished two tasks this week… adding more robustness to the movement grammar, and integrating speech recognition.

In particular, I added the abilities to parse commands such as:

• “go left/right” — turns left 90 degrees and then moves the given distance.
• “wait” — pauses the robot for x seconds.
• “stop” — stops all movement.
• “turn around” — turns 180 degrees, optionally “to the left”/”to the right”.
• “spin around” — turns 360 degrees, optionally “to the left”/”to the right”.
• “take a picture” — takes a photo using the fluke camera, which is transmitted to the computer via bluetooth and displayed.  This functionality is still a little buggy, which I suspect is due to some instability with my version of the Python Imaging Library.

I also loosened up the grammar in some other places to make it more flexible.  I wrote the following DCG predicate which has really come in handy:

optional(X) –> [X] ; [].

This is a case of something so obvious, that I don’t know why I didn’t think of it earlier.  That simple little line of Prolog code enables the insertion of “optional” words or phrases in a sentence and has become invaluable.  It is analogous to square-bracket notation in EBNF grammar definitions.

Integrating the Microsoft Speech Engine with the program proved to be quite a bit easier than I had imagined.  I was able to do this via a library called PySpeech.  After running through only one round of Microsoft’s voice training program, the engine is fairly accurate and fast enough to be useful.  So now, the parser can be run in either text or speech mode– pretty cool!  It is a blast to call out voice commands and see the scribbler react.

Next week I plan to wrap up the feature side of the project by adding a few more fun things (“beep” springs to mind) as well as conditional statements (“go forward until you hit a wall”).  All that will then remain is to flush out my comments and documentation.

This week, I worked solely on the Prolog natural language parsing component of the project.  This is basically a set of grammar rules along with very simple code-generation based on the results of the parse. I made quite a bit of progress this week… Basic movement commands are now parsed, along with quantities and conversions of a handful of different units. For example, you can now tell the Scribbler such things as:

* go forward for 3 seconds and then turn left
* drive forward three feet and then turn right 180 degrees
* move backwards 6.5 inches

And so on and so forth. Here is a sample run:

Enter a command: go forward two feet, turn right 90 degrees, drive backwards for three seconds
Prolog query executed: parseToCode([go,forward,two,feet,and,turn,right,90,degrees,and,drive,backwards,for,three,seconds], Status, CodeOut).
Status: valid
CodeOut: [‘forward(1.0, 4.1)’, ‘turnRight(1.0, 0.7875)’, ‘backward(1.0, 3)’]

A big component that I had to work on this week was english -> integer translation. It actually requires quite a bit of code to convert an English phrase like “nine hundred forty two point threee” to “942.3”!

My next task is to add some “fun” commands to the program; beeping, taking pictures, and maybe a wander/doodle capability.  I would also like to add some flavorful error messages to return upon a failed parse.

Another potential goal is to add conditional statements to the parser, i.e. parsing of a statement like “go forward until you hit a wall and then turn left and take a picture.”

Here is the source code for the Prolog parser as it stands today…

Read more

This week I finished the python script that handles initialization, user input, and all messaging to & from Prolog. All that is left to do is the natural language parser component in Prolog.

I was learning Python as I wrote this, so a lot of this code could be certainly written in a more elegant way. But it gets the job done.

The biggest obstacle was a silly grammatical problem I encountered was interpreting the Prolog list of code to be executed into Python code. A typical prolog code result arrives in list format [forward, ‘1.0’, ‘3.0’]. I was trying to parse that into Python as forward(1.0, 3.0). Notice the lack of string quotes around Prolog’s forward; this is because in Prolog, its data type is “atom” rather than “string”.  But this caused no end of trouble when pulling it into Python, as it was interpreting forward as a functor rather than as a string.  No problem, right?  It should be easy to just cast it as a string in Python.  But despite all my efforts, it wouldn’t work…  casting it was yielding a big mess, since it was interpreting it as a function pointer. I ended up having to resolve this issue on the Prolog side with a bit of a hack to get quotes around outgoing atoms.

Anyways, here is the Python code.  The main routine parses user input, passes it to Prolog, and then checks for a solution and returns a PrologResult instance with fields prologcall, status, and pythoncode.  Based on the result of status, it then executes pythoncode.

A typical command of “scribbler move forward” might yield the following PrologResult:

• prologCall = parseToCode([scribbler,move,forward],Status,CodeOut).
• status = valid
• pythoncode = [forward(3.0, 1.0)]

Read more

I came down with a brutal cold this week and did not accomplish as much as I would have liked to.  I did make some decent progress, however.

I spent an evening browsing the user documentation for the Myro library, which is a collection of Python functions to send basic commands to the Scribbler.  I also worked my way through “Learning Computing With Robots” by Deepak Kumar, a freely-available textbook about using Myro as a platform to learn Python.  In the process, I sharped my Python skills and learned my way around Myro.

My other, somewhat more significant advancement this week was successfully integrating Python, Myro and Prolog.  I found a hobbyist-made library called PySWIP to facilitate messaging between the Python shell and SWIPl, the implementation of Prolog that I am using.  Unfortunately getting this to work turned out to be a nightmare because of dependency and legacy issues… Myro runs best on Python 2.4, while PySWIP requires Python 2.7 and wouldn’t work with newer versions of SWIPL.  I had similar issues regarding to 32-bit version versus 64-bit versions.

I could not find any online to solve my quandary– I think I may be the first person to ever try using Prolog to control a Scribbler!  After brute forcing my way through several different combinations of versions, eventually I got Myro to run under Python 2.7 by updating all its dependencies manually, and found a version of SWIPL that was compatible with PySWIP.

So now I can import a Prolog file into Python by inputting:

> prolog.consult(‘sample.pl’)

And query it from Python by inputting:

> prolog.query(“male_ancestor(edward_vi, Ancestor)”)

Which returns a Python generator with all of the solutions that Prolog finds.  Pretty cool!