Since a while I am planning to build a second, more professional robot with additional skills. Everything is collected up to know

But configuring everything in hardware and software still needs time.

The Idea

This robot is meant to be a combination of a base robot, going on wheels, and a quadrocopter with gripper (mechanical/magnetic), realizing the arm of the robot. While the base robot already has an ASUS Xtion 3D sensor, as well as an IMU and a binocular camera system, the flying arm gets at most a simple camera sensor for tracking objects during flight.

The main tracking is supposed to happen by the 3D sensor, which is complemented by a stationary Microsoft Kinect 2 for higher resolution.

The whole setting is of course meant as an indoor research project (the 3D sensors do not work that well in the wild). The first big aim is to transport forks and other little things within a fixed apartment, combining the flying quadrocopter (with magnetic gripper) and the driving base.

Impressions

Quadrocopter, Wild Thumper 4WD (with gripper, which is going to be part of the quadrocopter), and an electro permanent magnet (gripper alternative for quadrocopter).

theano_models is a high-level package around theano which implements a Model type to pre-structure the (occasionally rather big) theano expression graphs. It is meant for clean programming of symbolic models, for instance Multilayer Perceptrons (MLP) or probabilistic modeling.

Its Strength

The Model gives a powerful view on the graphs, ready for interactive coding / prototyping, while being transparent and lightweight. It strengthens the underlying simplicity of theano.

For those, who know about the OpFromGraph by theano, this is a generic, faster, and better integrable alternative, at least I think so. It offers maximum control, while still making higher-level pre-structuring as simple as working with theano expressions directly.

Preview

It was developed as part of my master thesis, which is not done yet, and hence the code is still private. It will be open sourced afterwards, extending also this little introduction.

As a preview, please consider this visualization of a MLP Model, consisting out of three layers (AffineNonLinear Model), which is fed into a GaussianNoise Model. You see / interact with the loss function, which is build from the probability distribution function of the GaussianNoise, in log space (for numeric reasons).

Try to click on the yellow nodes:

The code will follow as soon at it is open source.

This module is a collection of helper functionalities which appeared to be useful here and there when working with Python.

I can especially recommend myfunctools which gathers really useful higher order functions, importantly myfunctools.fmap, myfunctools.compose and others. Also, take a look at myobjects, especially myobjects.Structure.

docear_helpers currently encompass the following utilities. I especially recommend init and autorename.

• init : initialize the current directory as a docear project. A references.bib file is created if not existent, where all latex references are stored.

• move : move / rename docear projects to another directory

• autorename : rename all pdf files according to their meta-data in references.bib (concerns all pdfs listed in the given docear project). This means that you can rename a “id034789234.pdf” file to “AuthorName(Year)PaperTitle.pdf” by one command, without breaking anything

• open_user_settings : opens the user.settings file in the default text editor

pyparsing_regex is a reimplementation of the pyparsing interface, building upon regular expressions.

It is intended to be a replacement for pyparsing where no recursion is needed, in order to speed up parsing significantly (factor 100).

Additions to Pyparsing

In addition to pyparsing, pyparsing_regex has a GroupLiftKeys class which works like Group, only that all string-keys are also available at the upper level (encompassing everything which belongs to that key further down). In a normal Group, the nested keys would be hidden from the upper layer, which might be unwanted.

Further a standard, efficient Repeat class is supported, which repeats an element an arbitrary number of times. It is the interface to a regex like (group-content){min, max}

Further References

See also http://pyparsing.wikispaces.com/

For a version of pyparsing using OrderedDict instead of dict (here and there useful to have) see https://github.com/schlichtanders/pyparsing-2.0.3-OrderedDict

After an internship at the robotics department of the University of Osnabrück a good friend of mine and I were motivated to build our own robot.

A Raspberry Pi was the favourite choice for controlling the system, as we wanted to use full-fleshed ROS in the end, the major Robot Operating System.

It took a while to setup everything, but it runs and is controllable with the keyboard. OK, that is not really that impressive, but for us it was a big success already.

Extra Features

Binocular vision was implemented, and also a 3D sensor, an ASUS Xtion, was bought. For mobile computation power, I configured my Nexus 4 to run Ubuntu and ROS, which indeed worked out. Of course, ROS also runs on the laptop’s Ubuntu, but it is great to know that in principle smart phones may be enough.

To Be Continued

For my Masters I changed the town and so this project went to sleep a bit earlier than expected.