Category Archives: Software

Delta Robot Update

Since I’ve last posted I’ve been up to a number of things with the delta robot. To start with here’s a video of the robot drawing circles:

Since that video was recorded the control electronics have been rebuilt to use an stm32f4 discovery board. The extra processing power this board provided enabled the inverse kinematics to be run on the delta robot rather than on the computer. This smoothed out the movement of the robot and simplified the code to control it on the computer side. Hopefully this will eventually allow me to integrate it nicely into 3rd party 3D printer control programs.

Another area I’ve been playing with is using the delta robot to draw animations using light and long exposure photography. The following video shows the first test I did.

To achieve this a camera running chdk was triggered from the computer using an arduino to capture a long exposure image. As the image was being captured the robot plotted out a 3D shape with an LED attached to its end actuator.
The whole process was scripted using python and repeated multiple times to capture multiple frames.
All the images were then stitched into a single video using ffmpeg.

Here’s a more elaborate animation made using the above technique.

Linear Delta Robot Inverse Kinematics Software Prototype

My previous post was showing off the initial OpenSCAD model of a liner delta robot project I’ve started. I’ve been interested in building a delta robot in some time but have been put off by the inverse kinematics, thinking it would be an absolute nightmare to figure out. Turns out it’s surprisingly simple for vertical linear delta robots and I’ve written a quick Python+PyGame GUI to work out the required Z positions for the 3 corners of the robot.

Liner Delta Robot Inverse Kinematics Software

The blue bars on the left show the required Z positions to reach the XY co-ordinate shown by the diagram in the center of the screen. The grey bar next to the individual Z positions shows the total Z height for the head of the delta robot. The dark grey crosses in the center region show the possible locations reached by the robot.
Setting the delta robots head XY position is done by clicking on a location. The Z height can be set by scrolling the mouse wheel.

For the time being I’m going to stick with pygame for the GUI components, I’m quite liking the minimal look. I may at a later date port it over to using QT.

Ultimately the goal is to use this software for testing the inverse kinematics model, as well as for debugging the delta robot during construction by implementing an interface to directly control the robot via USB. At a later date G-Code interpretation may be added.

This first version of the software is not in a fit state to release. Once I’ve cleaned up the code I’ll stick it up on GitHub under a GPL license.

Introducing undleBundle!

Tired of looking up the commands to compress/uncompress files via the command line in Ubuntu? Don’t know your -zcvf’s from your -zxvf’s? Then undleBundle may be just your cup of tea!
undleBundle is a python script which wraps a bunch of command line compression/decompressions tools together in simple interface with sensible defaults. Want to make a tar.gz? simple, just run:

bundle filename.txt

want to uncompress your file?

undle filename.tar.gz

want to select a specific compression type to use? (assuming you have the correct tools installed it currently supports tar.gz, rar, zip, tar and 7z, but adding new types is a breeze) Use the -t switch (only used in bundle, undle will automatically determine filetype for decompression)

bundle -t zip filename.txt

want to also select a new filename for the created file? Also simple!(see, a pattern is emerging) use the -f switch:

bundle -t zip -f filename.txt

And that’s it, that’s all there is to undleBundle.

Python – OpenCV IplImage To PyQt QImage

This is a quick post which will hopefully save someone some time. I spent far too long trying to figure this out.
If you’re using Qt and OpenCV with python and want to show an opencv iplimage within a Qt Widget here’s a quick hack to convert the iplimage to a qimage.

class IplQImage(QtGui.QImage):
"""A class for converting iplimages to qimages"""

     def __init__(self,iplimage):
         #Rough-n-ready but it works dammit
         alpha = cv.CreateMat(iplimage.height,iplimage.width, cv.CV_8UC1)
         rgba = cv.CreateMat(iplimage.height,iplimage.width, cv.CV_8UC4)
         cv.Set(rgba, (1,2,3,4))
         cv.MixChannels([iplimage, alpha],[rgba], [
         (0, 0), # rgba[0] -> bgr[2]
         (1, 1), # rgba[1] -> bgr[1]
         (2, 2), # rgba[2] -> bgr[0]
         (3, 3) # rgba[3] -> alpha[0]
         self.__imagedata = rgba.tostring()
         super(IplQImage,self).__init__(self.__imagedata, iplimage.width,iplimage.height, QtGui.QImage.Format_RGB32)

All in all it’s fairly straight forward, an example case is something like the following:

#Create a 3 channel RGB iplimage (could be from a webcam ect.)
iplimage = cv.CreateMat(iplimage.height,iplimage.width, cv.CV_8UC3)
#Turn it into a qimage
qimage = IplQImage(iplimage)

It works by sub-classing QImage and overloading the constructor with a new one that accepts an IplImage, this then has an extra alapha channel added (to make it compatible with the QImage pixel packing) and finally the __init__ method of the superclass (QImage) is called with the data from the iplimage passed into it as a string.

The important part is:

self.__imagedata = rgba.tostring()

This keeps a reference to the image data so it doesn’t go out of scope when the __init__ returns. (the QImage constructor that accepts image data doesn’t keep a local reference to the data, so you have to make sure it isn’t lost.[at least i think that’s right])

“Clothed all in the morning you know by the river” – A Lite Introduction to Markov Chains and Nonsense

(Poorly written and badly structured Python code for this blog post can be found here, this makes quite a fun and quick project if you’re trying to learn a new language)

Imagine reading the following half of a sentence in a sherlock holmes novel:

"My name is Sherlock".....

What is the probability the next word will be “Holmes”? Well, thanks to project gutenberg who host public domain e-books, and the wonders of “find” in nearly any text editing software, we can see that that collection of words appears exactly twice in all of the Sherlock Holmes books, and more importantly for us, 100% of the time it appears it will be followed by the word “Holmes.”

Now lets look at the following fragment of that previous sentence:

"name is"

We can probably assume the range of words that are likely to follow this fragment within the books is greater than just "Sherlock",and with a bit of tedious searching, we can find exactly what words appear and the number of times that they do.

From this table of figures it is now possible to derive a series of probabilties, so that now when you read the fragment "name is" in a Sherlock Holmes novel you can thoughtfully pause before reading the next word and take an informed guess at what it will be.

What if we now take the pair of words:

"My name"

What is the range of words that could follow that fragment within the books? Or even better, what if we take the entirety of the Sherlock Holmes novels, split it all down into consecutive pairs of words and, for every pair, calculate the probabilites for the next word? What we’d have would be a model we could use to predict the next word for any given pair of words within the books of Sherlock Holmes.

Lets abstract this idea a bit:

We have a model of a system, which given a state of the system, we can use to predict the next state.

In our case the “system” is the books, the “state” is a pair of words, and the “model” is a massive list of all the pairs of words in the books with a corresponding table for every pair showing the possible following words (the “next state”). An important thing to note is the predictions are based purely on the current state, no previous states are considered.

Pretty cool eh? but what to do with it? Well, there’s probably some pretty interesting analysis/classification problems this could help with, but I’m just going to use it to make nonsense sentences.

To make nonsense we first pick a random pair of consecutive words from our book as a starting point. We then look at all the possible following words in our model and roll a virtual weighted die to determine what the next word should be. We then take the 2nd word from out first pair and our new word, form them into a new pair and find the possible words that could follow this new pair….and so on.

The result of this can be quite interesting, sentences can seem to have a structure and a semblance of meaning, yet still make no sense. This is due in part to some pairs of words having very few possible words following them (for example "is Sherlock"), while other pairs of words have many possible words that could follow them (for example "it is").

Where can we go from here? Well, there are a number of possible options, instead of using whole words, why not use letters to make nonsense words? Change the number of consecutive words used to create the model and see what happens (my guess is you’ll reach a point where trying to get a random sentence will result in just reciting parts of the book word for word), how about feeding two different books into the model instead of just one? (feeding Sherlock Holmes and the Wizard of Oz in results in some pretty weird sentences that seem to “flow” between Oz and Sherlock).

Kinect & Qt

Impulsed bought a kinect and decided to write a Qt wrapper for it. A fork of the libfreenect git repo with the wrapper included is available here at my spangly new github repository (Not quite figured this whole git malarky out so it’s probably gonna get broken at some point).
Heres a screen shot of the output from the RGB and Depth cameras (note: depth has had it’s dynamic range reduced from 11 to 8bits).