I recently gave an update about my research program and the state of integration between language, vision, and cognition. There should be an update about robots sometime soon!
Anyone who enjoys functional programming is going to have a rough time on Android. You’re basically locked into Java because most of the functionality of the OS, the UI for example, isn’t exposed to C code. So it’s time to find a Scheme implementation that works on top of davlik. Turns out Kawa works pretty well and it’s blazing fast. Unfortunately there really isn’t enough information out there about how to use it, Kawa’s documentation leaves much to be desired and its error messages can be beyond cryptic.
I translated some examples from Java into Scheme and set up some scripts to help with setting up new projects, you can everything on github. At the bottom of the readme file I documented some of the more exciting error messages. Even though I tried to stay at close to the original Java as possible, even when it felt pretty disgusting, I couldn’t help myself in some cases. Take starting a new thread. In Java that takes several lines and a new class, and due to the lack of macros can’t really be shortened much. In scheme with a trivial macro this becomes simply
(thread code ...). It doesn’t get more beautiful than that. I’ll be putting up a full Android application sometime in the near future.
In the lab we recently bought two bleeding edge PowerEdge C6145s. Each has two servers with 48 2.5GHz cores. That’s a pretty big bang for your 2U. The danger with brand new hardware is that you also get brand new bugs and Dell, who was as confused as we were, didn’t let us down at all. One of the new boxes felt slow, real slow. Actually, per-core it was on par with my cellphone, 3.1 times slower than the other box. The machines are identical in every way. Turns out after lots of time wasted trying to figure out how the machines are different from each other, one of the power supplies had a yellow light and resetting it fixed the problem. But the strange part was that /proc/cpuinfo never showed any hint of frequency scaling even thought it must have been happening in the background anyway. More hardware, more problems…
If you’re thinking of buying a copy of the Complete National Geographic but are weary of having to run it under wine, despair no more. It’s written in Adobe Air and runs acceptably well. I used wine to install it so I could get at the air package inside, but it turns out NGM is nice and provided us with a handy download for the viewer. I couldn’t get it to run without unpacking it, something about an API version mismatch (originally I repacked it with the correct settings but that’s overkill). You can just unzip it in its own directory and run it with something like: /opt/AIR-SDK/bin/adl -nodebug /opt/AIR-apps/CNGViewer/META-INF/AIR/application.xml /opt/AIR-apps/CNGViewer &> /dev/null. Note that as with most Adobe technologies on Linux it’s needlessly CPU intensive, it pegs my CPU at 100% regardless of what frequency it’s set at.
As for the content itself, it’s fun; I specially like how many of the more interesting articles are organized into reading lists. My main complaint is that they used jpeg to compress the images and did so at really high compression leading to lots of artifacts. Looking at a page that has a black background can be painful. Which is a bit of a shame as one of the greatest things about NGM is the high quality of the photographs and printing. The viewer also has way too many needless animations and there’s sadly no way to disable them. So I wouldn’t mind knowing the file formats involved, an app for the n900 or more importantly a viewer that just does its job and gets out of the way would be great.
I had the great fortune of getting into the whole Linux thing and playing around with computers thing when I was really young. When you’re in 6th grade you’ve got unlimited amounts of time, I at least had a propensity to make things crash intentionally (I guess that’s why I’ve been connected to the mains three times). So of course once you get used to Linux the interesting question of sabotaging it in progressively more creative ways comes up. My housemate managed to surprise me today though, he removed the SD card that contained his running Android system. You’d rightly expect it to crash sooner or later but check out the screenshot; in the crashdump somehow the screen looks as if it’s trying to face the other way around. That’s a new and exciting feature. Click the picture to see it in all its glory.
When giving a talk about robots it would be a criminal offense not to include videos. Unfortunately the options for outputting pdf files from LaTeX with video on Linux are sort of slim. You can chain yourself to Acrobat, but even then it uses realplayer and I don’t know about you but my distribution got rid of it years ago. So you’re left with showing grainy and rather choppy flv files. You can try okular or evince but I’ve yet to find a single movie that they would play reliably without crashing. I also wouldn’t want to give a talk with evince as it’s sort of buggy, nothing like standing in front of a hundred people and it showing a blank screen. This leaves xpdf.
xpdf is sort of ugly being motif-based and all but it’s great for showing a fullscreen talk. Unfotunately since Beamer embeds videos as hyperref links you end up having to click the picture and press ok in a dialog box. Not fun and it sort of throws you off while giving a talk. I’ve made a version of xpdf that eliminates the dialog box (with a new flag -autoexec), automatically runs any external links when loading up a slide and puts the size and location of the link in environment variables so that mplayer can pick them up. There’s still a little flicker as it invokes mplayer to draw the window but I’ll handle that when I have a bit more time. Talks look and feel much slicker now. Although you really don’t want to use this as your default xpdf, running external programs embedded in a random file you download is the definition of a security hole.
And who knows? Perhaps this even will be usable on an overclocked n900.
My laptop has been keeping my room nice and toasty for the past few hours and now 240 chicken eggs are available from the repository! That’s out of my list of 290 eggs. A caveat, these are all entirely unmodified with dependencies automatically migrated into the deb. This means that the eggs depend on each other correctly but eggs can’t record dependencies on external packages; so you can install an egg and not have the appropriate deb package from the standard repositories. I don’t particularly want to go through 240 eggs and record this, so we’ll have to live with this situation for now. It shouldn’t affect many packages anyway.
In the process I’ve built the most inefficient package manager that I could, it finds the fixpoint of a function which filters out working packages from a list. Oh well, at least it’s short.
OpenGL is one of the reasons why many of the packages fail to build. Unfortunately the n900 only provides OpenGL ES which eliminates a lot of the fixed pipeline and changes almost entirely to a programmable one. That means a lot of the old OpenGL API was removed. Until someone comes out with an OpenGL layer for the n900 this this won’t be available.
srfi-19 doesn’t build, which knocks out most of the remaining packages. setup-api looks in the wrong place for the srfi-29 bundles, I’ll have to look into it at some point.
You can find a list of all the eggs and their statuses here.
The n900 really needs some sort of voice recognition. While the keyboard is nice, it’s sometimes a real pain and I really just want to be able to tell it ‘find all restaurants nearby’ or ‘call X”. There are a few options when it comes to voice recognition, I think the most likely to work well enough is CMU Sphinx especially now with the advent of pocketsphinx. pocketsphinx is specifically designed to be small and fast and it seems to deliver on both points. I build the pocketshpinx libraries for the n900 (patches to the upstream version of pocketsphinx are available here) and I’ve been playing around with them for the past two days. The accuracy is good enough and integration with the rest of the system so that you can give commands like switch task should be doable. To test things out and get my feet wet I’ve been using a quick python script called voximp, it’s simple to set up and it’ll let you issue commands immediately.
The new packages are available from the standard optified repository: gstreamer0.10-pocketsphinx, libpocketsphinx1, libpocketsphinx-dev, libsphinxbase1, libsphinxbase-dev, pocketsphinx-hmm-en-hub4wsj, pocketsphinx-hmm-en-tidigits, pocketsphinx-hmm-zh-tdt, pocketsphinx-lm-en-hub4, pocketsphinx-lm-zh-hans-gigatdt, pocketsphinx-lm-zh-hant-gigatdt, pocketsphinx-utils, python-pocketsphinx, python-pocketsphinx-dbg, python-sphinxbase, python-sphinxbase-dbg, sphinxbase-utils.
Hopefully I’ll get around to wrapping all of this in scheme bindings and a gui sometime soon, and now that I have MCE bindings I should even be able to implement placing calls by voice tags.
The main holdout to submitting the qt egg back into chicken have been the Maemo bits that crept in. Now all of that code has been moved into its own, chicken-maemo, egg so hopefully I’ll get around to submitting diffs against svn to the list later today.
With Maemo having its own egg I don’t feel bad extending adding in support for more of the basic infrastructure anymore. Now we have a full set of bindings to MCE over DBus, aside from the stackable windows, notes, banners, and orientation control that we had before. This means accelerometer data, calling, vibrator support, locking various aspects of the device, using the leds, and more. Sadly I still haven’t figured out how you generate documentation for eggs.
I’ve also built my first app that will see real-world usage. My advisor is a session chair ICRA and he wanted an application to manage speaker time during it (they’ve got this 12 (talk) + 2 (questions) +1 (cleanup) minute format). This also brought up an interesting problem of how device rotation is going to work when the guy needs to change its layout, I’ll have to investigate further. Now that the MCE bindings exist, I should add things like vibration and manging the screen state. Still, it’s not a bad app for an hour of work and it’s a decent demo so it lives in the chicken-n900 repository as test4. Obligatory screenshots follow.
I recently added a whole bunch of new eggs: check-errors, doctype, easyffi, eggdoc, filepath, fmt, foreigners, html-parser, htmlprag, input-parse, iset, json, matchable, message-digest, miscmacros, numbers, packrat, protobj, rss, setup-helper, sha2, silex, sql-null, sqlite3, ssax, sxml-transforms, sxpath, synch, test, utf8 and yelp. These are all unmodified and just running the debianize-egg script on them suffices so I didn’t need to import them into my code repository. Adding in this optified repository on your n900 and doing an apt-get chicken-eggname will fetch the egg and its dependencies. Hopefully all eggs will be available soon once I finish the autobuilder script.
chicken-qt got networking support recently, nothing fancy just HTTP Get for now. On the upside it integrates well with Qt so it’s asynchronous by connecting to something like the done() signal. Check out tests.scm for an example.
For the past two weeks or so I’ve been working on getting Chicken scheme up and running on my brand new Nokia n900. The n900 is really lacking a lot of apps and it’s time to have an environment to quickly develop them. Python isn’t an option, try out gPodder and see your device slow to a crawl and your battery drain, so Scheme it is. Things are looking rather good and in the process I’ve expanded the qt egg, hopefully it can leave it’s tech-demo status now.
– support for QObject attributes
– QVariantList and qvariant are supported so interfacing with qt is a
lot more natural
– removed the simple receiver and replaced it with an object that
crafts the correct slot on request
– arbitrary number of arguments to slots and support for some common
argument types (although notably pointers aren’t yet handled)
– invoking arbitrary qt methods/signals/slots
– support for qt-maemo (stackable windows and various information boxes)
– it now builds in C++ so that making bindings is faster
– qt-dbus is now supported so that polling isn’t required, it too
works with an arbitrary number of arguments and provides all of the
basic capabilities: receiving/sending dbus signals, method
invocation, and exposing methods over dbus
– 0% cpu usage when idle so it’s battery-friendly
– it’s debianized along with its dependencies so that it can be
deployed to the n900
– ported the dbus egg to version 4
The updated qt & dbus eggs along with debianized versions of their dependencies are here. There’s also a tiny script for quickly debianizing eggs with correct dependencies, perhaps someone is interested in reviving the debianization effort.
I’m new to low-level hacking with chicken, so hopefully I haven’t abused the ffi/gc in some horrendous/unsafe way.
I’ll try to keep http://0xab.com/n900 current with new progress.
Hopefully I’ll figure out how one documents eggs these days, split off maemo stuff into its own egg and get around to contributing these back to the main egg repository.
Now that I’m giving talks and presenting posters more often, it’s time to make them look as if they’re actually from Purdue. I’ve written up a a Purdue Beamer theme, you can download a tarball here. This contains both a poster theme and a theme for talks.
For posters you’ll have to edit the style file to change the attribution, url, and authors. Maybe someday I’ll make it take arguments. Other needed files, like the logo, along with an example, can be found here.
If you’re going to use either of these the best thing to do is to check out the git repository.
So it turns out I’m heading to ICRA 2010 soon. Our paper on robotics, vision and learning was accepted! We describe how we get the robot to learn board games by watching another robot play. Lots of pictures, videos and details are available on my research page.
More details about that work is on our website for the submission.
Andrei Barbu, Siddharth Narayanaswamy,
Jeffrey Mark Siskind,
“Learning Physically-Instantiated Game Play Through Visual Observation”,
Proceedings of 2010 IEEE International Conference on Robotics and Automation
More videos of the robots.
I’ve recently been editing my CV getting it into shape for summer internship applications. I think it’s starting to look decently good. The LaTeX theme (well, it’s just the resume itself, for now, maybe I’ll make it into a theme later) for it is now much easier to use, and anyone should be able to quickly write up a resume with it. No fiddling with LaTeX required. I’m told that I’ve sinned by breaking the cardinal rule of a one-page resume but oh well, there’s a lot of stuff to talk about. The only thing I might want to change is the fact that it maybe it’s missing some color.
A lot of people with Intel wireless cards seem to have issues connecting to hidden networks with NetworkManager. Unfortunately my university only provides wireless through a hidden network. Although of course there’s a handy one that broadcasts its SSID, doesn’t require a login and explains what the SSID of the hidden one is (how this is more secure is beyond me). The problem is that NetworkManager just gives the card the BSSID and it needs the SSID as well, even though it really shouldn’t. This patch applies to head as of the end of 2008 and fixes the issue. Some distributions patch their NetworkManager in order to include something similar to this. This bug has been around for as long as NetworkManager has existed from what I can see, I wonder why it hasn’t been fixed.
I recently bought a Gateway/MPC tablet (E-295C, equivalent of C-140-something). Whoever designs these should probably quit while they’re ahead. They’re one of the worst computers to open up that I’ve ever seen. While nominally they’re designed to allow easy access opening it is always a struggle between applying just enough force to unclip a piece while not breaking it. Although that isn’t the point here.
My tablet kept overheating, I had to set up a script to balance the CPU throttling vs the temperature. Not fun. Then there was another problem, the GPU. Any time it would come out of powersave mode, like when fglrx starts, the tablet would turn off seconds later. It was also responsible for many of the heating shutdowns the CPU would be at a cool 85 degrees Celsius and the BIOS would turn things off because the GPU had reached 125 degrees.
So I sat down yesterday and opened it up in order to see what’s going on, what I found was downright ridiculous. The first thing that struck me is that the heatpad, which covers the CPU and GPU, wasn’t tightened properly. This meant that when it was facing downwards, like in normal operation, it would hang creating a tiny gap between it and the GPU. But then it got worse. So I took the headpad off. Of course I found a CPU and a GPU with plenty of therm.. Yeah, right, I hope you didn’t think that’s where this story was going. Neither the CPU nor the GPU had any termal compound on them. The surface was extremely rough, I’ve never seen a heatsink of this low quality before. This is probably a good time to mention the scratch marks on the heatpad. So I got some Arctic Silver and put the headpad on correctly. The result? No more heating issues. Not only can I start the GPU now, I can run CPU-intensive and GPU-intensive tasks at the same time. Thankfully the CPU is ok with working at 93 degrees…
Fixing this motivated me to get the Windows driver for those pesky tablet screen buttons that weren’t working. An hour or so later in IDA and now there’s a Linux daemon that will execute commands on button presses. If you have a Gateway/MPC tablet and want button support you can get it here. I much later found out about a kernel module that does the same but sadly it’s never worked for me.
Note: Sometime after this work better bindings have been submitted to Hackage, so this is no longer relevant.
I decided to make Haskell bindings for LLVM today. The build system is a horrible mess so things don’t actually install, they just build in-place. Mandatory brainfuck compiler included, it assumes a 64bit architecture but I’ll clean that up eventually. Just drop into bindings/haskell and run make. LLVM is fun to write code with.