CUDA 3.2 has been released, with H.264, random number, and sparse matrix support, along with the usual smattering of improvements.
Intel OpenCL has finally been publicly released as an alpha. Ofcourse it only targets the Streaming SIMD Extensions (SSE) for Intel CPU's. Still no word from IBM or ARM on their OpenCL implementations.
General Electric have released an embedded GPU platform. This looks great, I don't know why more people haven't mentioned this. GPU's are perfect for many of the processing problems we deal with for autonomous vehicles in interpreting vision, RADAR and LIDAR data.
The Microsoft Kinect is a very cheap 3d vision system that provides performance somewhere between stereo cameras and more expensive time-of-flight (eg:PMD) cameras. It was actually something I was hoping would come out before the end of the MAGIC 2010 competition, but unfortunately the release date was scheduled for after the end of the competition.
The device was "hacked" almost immediately after its release and it has already been widely adopted, I think this will do wonders for the robotics/computer vision/stereo-vision community. So far all we have seen are relatively simple applications, but I'm sure people will start pushing the interpretation of the data to its limits soon, especially given that game-developers are now going to be playing with this kind of data. The Wii has already done fantastic things to push the knowledge of filtering and dealing with noisy sensor data to a wider audience, and I expect the Kinect will have a similar effect.
The OBJ importer in blender actually supports batch-import, to enable this you need to hold SHIFT when you select the importer from the File->Import->Wavefront (obj) menu. You will then be prompted to select a directory. Unfortunately, the importer will import each OBJ file into a separate (and incorrectly constructed) scene. To alter this you will need to edit the "import_obj.py" script and comment lines 1164 and 1165:
for f in files:
#scn= bpy.data.scenes.new( stripExt(f) )
#scn.makeCurrent()
Saves you from having to write a batch import script yourself! A great utility to convert to OBJ files from various formats (with a batch-convert mode!) is polytrans.
All of our teams hard work has paid off as we claimed 4th place in the MAGIC 2010 competition! Our system was actually very close to what the top two teams had implemented, and with a bit of luck we might have been able to snatch a prize. Maybe next time!
Above is a photo composition by Rob Reid of all the robots at the challenge stand, ordered from 1st to last (Left to Right). Team Michigan came in 1st, followed by University of Pennsylvania, and RASR took 3rd place. It's been a great opportunity to talk to all the other teams and see how they solved their problems and to see how many of the same headaches we all had. The LWC was a great place to meet and greet a number of industry people and also a nice crowd from DSTO who showed a lot of interest in what we have done. It was great to get so much positive feedback from everyone on our system.
The Ram Shed challenge was a media-event for the MAGIC 2010 competition. It was our first chance to see all the other teams robots and a chance to see them all at work. Our robots performed very well one the day and navigated the maze better than any other team, the entire open section of the maze was explored fully autonomously requiring no human interaction at any point. All of this despite the fact our wifi gear was mounted outside and failed during the rain, after a few minutes we were able to replace it and continue our mapping operations!
The photos show our robots at the starting line, followed by them moving off in the starting arena, then navigating the maze, and finally the prize.
A fabulous effort by our team, and I'm extremely proud of our achievements on this day!
Posts
