Wishlist

3D user-interface (open)
3D-Mapping (open
Joint Rescue Forces list of tasks
small tasks
Reconnect when in range
Use status message to select view.

Started Labbook 2019.

December 2, 2018

The multi-robot rendezvous paper has an accomponying dataset

November 28, 2018

Scanned Mathworks presentation on Nvidia conference, where they demonstrated LIDAR point-cloud clustering, trained on a CNN (lidarlinknet) and than converted to native GPU code. Should look if I can reproduce that demo. The dataset (velodyne) seems to be here.

November 22, 2018

Received an invitation for a chapter on a book of Rescue Robots from a publisher on Predatory publishers

November 14, 2018

Promissing DCOP algorithms are DSA-SDP and GBDA. Should check those algoritms on RMAS-BENCH.

November 9, 2018

AutoML has a number of basic AI techniques implemented in Java, which can be combined with a software configurator (including a hierarchical planner to optimize the configuration). i

November 1, 2018

This paper is using Matlab's MRSim for SLAM in a search and rescue scenario.

October 4, 2018

CPLEX is IBM's C-implementation of the Simplex algorithm, but this is developed into a full studio of Linear Programming and Constrained Programming. It consists of studio with algorithms in C and a Python port (including Jupyter notebook tutorial). It would be interesting to see what can be done here with Rmasbench problem. According to background article, Mixed Integer Programming Problems are very suitable for Vehicle Routing.
The suite also has a Java API.
Read two articles about MIP, one about Machine Learning, another on the history of MIP. The latter article showed that CPLEX improved every version a factor 3x.

September 29, 2018

The article from Nobuhiro Ito can be found at dblp, but still not imported at robocup.org. Paper was already added to Rescue Simulation Publications.

August 28, 2018

Cooperate A* as described on AAMAS Multi-Agent Pathplanning workshop would be great for the December workshop.

August 23, 2018

Downloaded from file exchange the example code of Deep Learning in Action - part 1. Actually, it are 4 demos. The third one (counting pills), contains a bug: I had to move the segmentation-code to the end (when a BW-image existed).
The fourth example contains a training with a GPU. With 'a zippy NVIDIA P100 GPU in roughly 20 minutes. Test images set aside give an accuracy of roughly 90%.'. The Titan Xp (eGPU) needed the same time (10/30 episodes in 7 minutes). In my case both the CPU (100%) and the GPU (97%) were fully used. The Titan Xp used 24 minutes 4 sec to reach 89.96% accurracy.

August 20, 2018

Last Thursday the DeepLearning demo was nearly finished, when I put him on pause. Nicely unmounted the GPU. The other applications handled this nicely, but MATLAB reports that it crashed to a low-level graphics error. To prevent another crash in this session, MATLAB is using software OpenGL instead of using your graphics hardware.
Reconnected to the GPU. Restarted demo after restarting MATLAB. After 80 minutes epoch 37/120 is already reached (compare this to the 3 days of June 30). Yet, the machine seems frozen.
Rerun, finally have the full demo. Network was trained in 262 min 16. Results were stored in PreTrainedCnn.mat.

August 16, 2018

Now Matlab 2018b recognizes my Titan Xp GPU, it performs gpuDeviceCount and gives ComputeCapability 6.1 at gpuDevice. Started DeepLearning demo which also worked on July 3 (but with a borrowed Titan Xp).
In 10 minutes it is already working on Epoch 5 of 120. Should be finished in 4h.
nvidia-smi should be installed together with the display-drivers. Yet, I have received 398.36 via WindowsUpdate, installing CUDA9.2 (again) would downgrade my drivers.

August 15, 2018

Could only find a drcsim installation at Wayback machine. Yet, drsim (v5.0.0) is still part of the 64bit Trusty packages (Ubuntu 14.04).
Under Xenial (Ubuntu 16.04) there are packages for servicesim-competition, robocup3d, ariac (Agile Robotics for Industrial Automation Competition), Service Academics Swarm Challenge.
Also under Vivid drcsim is already gone from the packages.
drcsim had dependencies on: libgazebo4-dev, ros-indigo-std-msgs, ros-indigo-common-msgs, ros-indigo-image-common, ros-indigo-geometry, ros-indigo-controller-manager, ros-indigo-geometry-experimental, ros-indigo-joint-state-publisher, ros-indigo-image-transport-plugins, ros-indigo-actionlib, ros-indigo-cmake-modules, ros-indigo-control-msgs, ros-indigo-image-transport, ros-indigo-nav-msgs, ros-indigo-nodelet-topic-tools, ros-indigo-robot-state-publisher, ros-indigo-ros, ros-indigo-rosbuild, ros-indigo-tf, ros-indigo-trajectory-msgs, ros-indigo-xacro, ros-indigo-gazebo4-ros, ros-indigo-gazebo4-plugins, ros-indigo-gazebo4-msgs, ros-indigo-laser-assembler, osrf-common (>= 1.1.0-3), sandia-hand.
This Korean wiki has a recent turorial on how to use the AtlasSimInterface in Gazebo.
Same site has instructions how to install drcsim under Kinetic. Kinetic combines with Xenial (Ubuntu 16.04).

August 8, 2018

Received the Nvidia Titan Xp. Mounted it in the Razer. Matlab gpuDeviceCount detects the device.
Tried to reproduce the DeepLearning demo3 on Ubuntu 18.04 with Matlab 2018b. Had the same issues as before with the dataset, but luckily I had a labbook.
Matlab reacts very slowly (even on simple cd), maybe I should reboot.
Rebooted, but I get from Matlab a gpuDeviceCount of 0. Thunderbolt also gave a warning, that a device was not working correctly. DeviceManager showed that the drivers where not correctly loaded. Uninstalling (and automatic reinstalling) solved this issue; my Desktop was on the monitor. Yet, still the gpuDeviceCount remained 0.
Rebooting to Linux again; also this reboot is really slow.
Reboot completely failed, asked for a fchk on drive.
Recovery ubuntu boot, selecting fchk followed by resume worked.
In Matlab2018b, trainingNetwork failed on ComputeCapability. Followed instructions of Matlab GPU computing, but gpuDevice didn't work because class parallel was not known. In Matlab2018a this worked, and Cuda version 9.2 and ComputeCapability 6.1 was indicated. Yet, here the mlx files couldn't be read (see Live Editor requirements). Just running the Demo1 *.m failed readData in trainNetwork. Running Demo2 failed on required AddOn GoogleNet (and AddOn Manager doesn't respond). Also manually downloading GoogLeNet from File Exchange didn't work for Matlab2018a.
Will try gpu benchmark example on Matlab2018a. That works, for Doubles a speed up of 4.7x is measured, the speedup of Single is a function of the matrix size, for the largest matrices the speedup is 20.6x.
sizeSingle: [1024 5120 9216 13312 17408 21504 25600 29696 33792] gflopsSingleCPU: [58.0327 205.5579 247.7993 197.1065 187.0655 190.6386 193.3857 195.5882 196.9357] gflopsSingleGPU: [101.9470 956.3970 1.6716e+03 2.1256e+03 3.0127e+03 2.4577e+03 2.8777e+03 3.4064e+03 4.0693e+03] sizeDouble: [1024 5120 9216 13312 17408 21504] gflopsDoubleCPU: [31.7532 72.3420 81.6168 85.6623 87.5513 89.2304] gflopsDoubleGPU: [24.8494 343.1366 383.1151 402.6244 414.3873 417.7095] speedupSingle = 1.7567 4.6527 6.7457 10.7838 16.1052 12.8922 14.8804 17.4161 20.6633 speedupDouble = 0.7826 4.7433 4.6941 4.7001 4.7331 4.6812 Strange enough the plot-function doesn't work in Matlab2018a. Saved the workspace and loaded it in Matlab2018b, and the figure was automatically drawn:
Not only Matlab works, also nvidia-smi nicely shows that the fan is running on 23% (while no process is active).
Copied /usr/local/cuda-9.2/samples to my home-directory. Made and run 1_Utilities/deviceQuery succesfully.
Made Mandelbrot, but it fails on:
CUDA error at Mandelbrot.cpp:971 code=30(cudaErrorUnknown) "cudaGraphicsGLRegisterBuffer(&cuda_pbo_resource, gl_PBO, cudaGraphicsMapFlagsWriteDiscard)"
Made simpleGL and simpleTexture3D, but fail on:
CUDA error at simpleGL.cu:422 code=30(cudaErrorUnknown) "cudaGraphicsGLRegisterBuffer(vbo_res, *vbo, vbo_res_flags)" CUDA error at simpleGL.cu:434 code=33(cudaErrorInvalidResourceHandle) "cudaGraphicsUnregisterResource(vbo_res)"
Imaging/recursiveGaussian fails for the Titan:
CUDA error at recursiveGaussian.cpp:312 code=30(cudaErrorUnknown) "cudaGLRegisterBufferObject(pbo)" CUDA error at recursiveGaussian.cpp:197 code=17(cudaErrorInvalidDevicePointer) "cudaFree(d_img)"
5_Simulations/nbody fails on:
CUDA error at bodysystemcuda_impl.h:183 code=30(cudaErrorUnknown) "cudaGraphicsGLRegisterBuffer(&m_pGRes[i], m_pbo[i], cudaGraphicsMapFlagsNone)"
Razer has now also the Core X, which allows cards that consume 650W.

August 7, 2018

Interesting 2005 paper on ICP with Sonar.

August 6, 2018

With a lot of effort, I was able to have CTR-ALT-F1 control again on my ubuntu partition of nb-dual.

July 13, 2018

Followed the instructions to install Bumblebee with command apt install bumblebee bumblebee-nvidia. update-alternatives --config glx did not work, update-alternatives --config x86_64-linux-gnu_gl_conf still did. No bumblebee option. Restarted, but login / CTR-Fn-F1 gave no reaction anymore.

July 13, 2018

In principle Thunderbolt 2 interfaces can be found on MacBooks only. The only exceptions are the HP Spectre 13-v100nb and the HP ZBook 15 G2 (two types), according to tweakers.net.

July 12, 2018

Should try the Bumblebee tricks from the advanced Linux eGPU guide.

July 10, 2018

Trying to get the external GPU working on my Ubuntu partition on nb-dual.
dmesg | tail sees a lot of thunderbolt messages (2 PCIe, onde DP/HMDMI), but nothing from Tesla or nvidia.
Following the instructions from CUDA toolkit.
Yet, lspci | grep -i nvidia shows nothing. Following the steps of XPS 13 9360, I see that cat /sys/bus/thunderbolt/devices/0-103/authorized gives 0, but that even as root I cannot overwrite this setting.
The package thunderbold-tools is not available for 16.04 Xenial, but Intel's thunderbold-software-user-space should work for 16.04.
Build and installed this software, and authorized the two thunderbolt with command sudo tbtadm approve-all. Now my mouse works! lspci | grep -i nvidia
3c:00.0 VGA compatible controller: NVIDIA Corporation GF100GL [Tesla C2050 / C2070] (rev a3) 3c:00.1 Audio device: NVIDIA Corporation GF100 High Definition Audio Controller (rev a1) The display shows not content yet.
Downloading Tesla drivers from Nvidia, with the C2050 in supported products. This forum suggested another route, but the suggested versions (390.48) was ahead of what nvidia suggested for the Tesla (390.30).
After installing the suggested drivers, the login crashes. Looked at the suggestions from askubuntu. Important is the trick CTR+Alt_F1 to login without graphics. My problem was not with .Xauthority, I am still the owner. I continued with update-alternatives --config x86_64-linux-gnu_gl_conf and selected /usr/lib/nvidia-390/ld.so.conf and restarted service lightdm start. I get my desktop, although not full resolution. The selection is still /usr/lib/x86_64-linux-gnu/mesa/ld.so.conf.
Connected the eGPU, and lsmod | grep nvidia showed nvidia drms. Switched selection again to nvidia-390. Got the following warning:
update-alternatives: using /usr/lib/nvidia-390/ld.so.conf to provide /etc/ld.so.conf.d/x86_64-linux-gnu_GL.conf (x86_64-linux-gnu_gl_conf) in manual mode
update-alternatives: warning: skip creation of /usr/share/grub-gfxpayload-lists/blacklist/10_proprietary-graphics-drivers because associated file /usr/share/nvidia-390/nvidia-390.grub-gfxpayload (of link group x86_64-linux-gnu_gl_conf) doesn't exist
Problem seems to be No discrete VGA device found.
Yet, now I can login again. An internal error indicated Xorg crashed in RRProviderAutoConfigGPUScreen.
The cuda-installation-guide was incomplete, installing cuda can be done with command sudo apt install nvidia-cuda-toolkit. Strange enough, this installs nvcc v7.5. In addition, nvidia-smi is not known as command, while it should be in package nvidia-390. Running /usr/bin/nvidia-detector returns 0.
cat /proc/driver/nvidia/version shows thaat v390.67 os active. According to the post-install instruction, there should be a /usr/local/cuda directory, but I couldn't find that.
Downloaded Cuda 8 again, but this time as runfile (which should create a NVIDIA_CUDA_Samples directory in my Home-directory (and a /usr/local/cuda-8.0). This will also install an old driver (v375.20). Installing the driver fails because X is running. Running the installation again, saying no on the driver.
Made DeviceQuery, but get CUDA driver is unsufficient for CUDA runtime.
Try the same, but now for CUDA 7.5 (Ubuntu 15.04 runtime). Here the suggestion is driver v352.39; I say no.
Installation fails on compiler. According to stackoverflow, As of the CUDA 7.5 release, gcc 4.8 is fully supported, with 4.9 support on Ubuntu 14.04As of the CUDA 7.5 release, gcc 4.8 is fully supported, with 4.9 support on Ubuntu 14.04. So installed g++-4.8.
That work around worked, but still the same CUDA driver is unsufficient. Two possibilities. Install the old v352.39 driver, or install CUDA 9.2. I will first try v352.39.
Even with lightdm stopped, the installation complained that a Xserver was running. Using the trick from stackoverflow and executing the following script solved this issue. Now I get another error-message: No CUDA-capable device detected.
if ! xset q &>/dev/null; then echo "No X server at \$DISPLAY [$DISPLAY]" >&2 exit 1 fi
Executing service lightdm start gives bbswitch: No discrete VGA device found, so bumblebee is active on my machine. Removed bumblebee, but still I have trouble to login again. Should do the trick with update-alternative again.

July 9, 2018

Explictlty setting VCBinDir to C:\Programs\Microsoft\VisualStudio\2017\Enterprise\VC\Tools\MSVC\14.14.26428\bin\HostX86\x64 in CUDA 7.5.props solved the problem with nvcc.exe. Raising the MSC_VER in host_config.h also works. Remaining is a missing helper_cuda.h.
The missing helper_cuda.h was due because I didn't copied the full CUDA Samples directory, adding the common-directory solved this. Now I get many errors due to the version of VC (v141).
So, also installed Visual Studio 2015. Looks like most of the VC-stuff in installed in C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC. Yet, was not able to retarget the solution to 2013 (v120). Yet, the 7.5.props is expected in C:\Program Files (x86)\MSBuild\Microsoft.Cpp\v4.0\V140\
So, downloaded Visual Studio 2013 from visualstudio.com.
With Visual Studio 2013 finally the nbody runs, it finds 1 device fro Simulation, identifies it as a Tesla C2050, but fails in bodysystemcuda_impl.h:160 code=46 (cudaErrorDevicesUnavailable). Running the program with argument nbody -cpu finally starts the program (yet not on the GPU):
According to this post, this is a particular problem of mixing OpenGL context with CUDA context. Tested vectorAdd (0_Simple), and that program works.
Next tried 1_Utilities\deviceQuery, which shows that the Tesla has 448 CUDA Cores running at 1.15 GHz.
Next tried 1_Utilities\bandwidthTest, which shows that initially the bandwidth is 1228 MB/s, which increases to 102186 MB/s in the 3rd pass.
Yet, 2_Graphics\Mandelbrot fails (also on the OpenGL context).
Tried 2_Graphics\simpleD3D9, but d3dx9.h was not found. Same for d3dx11.h
Next 2_Graphics\simpleTexture3D, but this program also fail on OpenGL context.
Next 3_Imaging\recursiveGaussian. This nicely reports: Tesla card detected, no OpenGL display, running the test in benchmark mode. Result of the benchmark is 2.22 Mpixels/sec.
Updated eGPU guide with a section on Windows Software.
Maybe a nice idea to look if dragonfly algorithm is also applicable in the agent-competition.

July 8, 2018

Installed CUDA v7.5 next to v9.2. Visual Studio can load the VS2013.sln, only retargeting fails because I have no wrrite permission on C:\ProgramData.
With a copy of the sample directory in C:\tmp retargeting works, only the nvcc.exe cannot find the compiler (specified with ccbin). The compiler is at C:\Programs\Microsoft\VisualStudio\2017\Enterprise\VC\Tools\MSVC\14.14.26428\bin\HostX86\x64, while v7.5 expects it at C:\Programs\Microsoft\VisualStudio\2017\Enterprise\VC\bin\x86_amd64. This should be corrected in C:\Programs\Microsoft\VisualStudio\2017\Enterprise\Common7\IDE\VC\VCTargets\BuildCustomizations\CUDA 7.5.targets (line 604).
Actually, the ccbin is defined in CUDA 7.5.props, where it depends on VCInstallDir. Copied the directort where nvcc.exe expects, but now it complains that the cl.exe specified in -ccbin is different from PATH. Looked in Powershell to echo $Env:Path. Quite a mess, with two CUDA bins in my path. Yet, no cl.exe in my PATH.
Opened a VC2017 developer commandline, and checked echo %Path% and echo $VCInstallDir. The later one is C:\Programs\Microsoft\VisualStudio\2017\Enterprise\VC\.

July 6, 2018

Even when the Tesla is connected, nbody reports no GPU found. Should look if CUDA has another check.
Tried if Matlab would start the training, but with ExecutionEnvironment=gpu it reports that it needs a GPU with at least capacity > 3.0. The computeCapacity reports that the drivers support CUDA 7.5. I have deleted GeForceExperience, so I cannot use that tool to update the driver. Just running the installer still fails.
I removed the Nvidia-directory, so installation of the Titan tools is no longer there, including the monitor nvidia-smi.
Douwe was also not able to run the nbody sample on one of the MLMs. nbody -compare gave an OK, the later calls just an error (device busy?!).
Looked how Matlab checks the number of devices. Most of the functions of distcomp are Matlab-functions, the underlying p-code can be found in directory C:\Programs\MATLAB\R2018b\toolbox\distcomp\gpu\+parallel\+internal\+gpu.
nbody does first a cudaGetDeviceCount. Commented out the exit, now it fails later on cudaGetDeviceCount with cudaErrorInsufficientDriver (code=35). Should try CUDA7.

July 5, 2018

Positive news about my laptop: the latest 2018 XPS 13 9370 now provides 4 PCIe lanes for its dual Thunderbolt 3 ports.
GeoForce experience recognized that the Tesla C2050 was connected and was trying to download the Tesla drivers (v397.44). Strange enough it couldn't connect to nvidia server.
Downloaded VisualStudio 2017 from datanose by following the To store at the bottom.
Building the initial sample (5_Simulations/nbody) failed because Cuda has to be installed after VisualStudio (CUDA 9.2.props was missing).
Seems that I have to versions of CUDA 9.2, one in Programs and one in Program Files. The one in Programs is the latest.
The installer removed everything in Programs. Tried again :-)
Major problems, with crashing setups, even from Cuda 8.
Solved it with installing CUDA with the GPU not connected.
Still, problems to start the startup sample. First the samples are quite well in hidden in C:\ProgramData\NVIDIA Corporation\CUDA Samples\v9.2\, second CUDA expected Windows SDK 10.0.1734, which is not installed by default. Only SDK 10.0.15063 and a 16... version were available. Next problem: my version of Visual Studio is too new: _MSC_VER 1914.
Installed the VC++ 2015 toolset (part of the Visual Studio 2017 installer). Still same problem, which seems to originate from the argument -ccbin "C:\Programs\Microsoft\VisualStudio\2017\Enterprise\VC\Tools\MSVC\14.14.26428\bin\HostX86\x64" to nvcc.exe.
From this Microsoft post, I can deduct that with version 4.14 _MSC_VER should be 1914 (and host_config.h goes to 1913).
Modified host_config.h (> 1913 -> > 1914). Now it works. Unfortunatelly, the eGPU is not connected, so I have to test tomorrow. Even testing nbody.exe -cpu from C:\ProgramData\NVIDIA Corporation\CUDA Samples\v9.2\bin\win64\Release\ doesn't work, even with this argument a GPU is requested.

July 3, 2018

According to this post, the situation with eGPU gets more complicated with the new TI83 Thunderbolt 3 chipset, because the laptops have to be officially certified.
Received an Titan Xp. That makes a difference.
First the drivers are installed, the capability detected is 6.1,
Most important: 1 epoch in 2m13. That would mean a full training (120 epochs in 4h).

July 2, 2018

Received an 8y old Tesla GPU (C2050) to experiment with in the Razor eGPU.
Preparing for connecting the eGPU, I have dowloaded the CUDA 9.2.88 toolkit, assuming that this also include the latest GPU drivers.
Also downloaded the Tesla drivers (v397.44).
After that, several frameworks are available, as indicated on deep learning frameworks page.
Setup of the eGPU was easy, just putting the GPU in. The Thunderbolt cable is rather short (30cm), it just fits on my desk. With the Thunderbolt cable connected, Windows directly recognized it as Thunderbolt device and asked for permission to connect (selected always). Windows installed several VGA and PCI devices (and requested a restart) Installing the Tesla drivers failed, because no compatible card was recognized.
Matlab seems to recognize the card, because the training seems much faster (first 16 iterations in 13 minutes).
Yet, when I ask in Matlab count=gpuDeviceCount, I get 0. Later I found because the card didn't get enough power.
According to this Matlab answer, this is possible when the CUDA driver can interact with the card.
According to akitio, not all computers support external GPUs. When checking Thunderbolt settings, my machine indicates external GPUs supported: No. I have the latest Thunderbirt drivers (17.1.64.250) and BIOS. Looked in my BIOS (1.4.0), but all Thunderbolt options are ready to go. Should check with a boot when the Razor Core is connected to my machine.
Added an external screen to the Razor, but my machine started up under Ubuntu. Followed the steps of eGPUs under Linux Guide, but update-alternatives --config glx says that there are no alternatives to glx.
Checked more /sys/bus/thunderbolt/devices/0-103/device_name, and device 103 is the Razor Core. The firmware has v26.0.
According to this post, it should work under Windows once I can get the nvidia drivers installed.
Still, no success. Also tried another Nvidia package, but to no avail. It could be that the GPU card itself is not fully inserted. Checked the AudioVisual service, but they don't have Thunderbolt2 to 3 converters (needed for Sander's Mac). Toto will bring his new laptop tomorrow. According to this post, Dell only supports eGPU over Thunderbolt3 on its Alienware machines. According to this post, Dell only supports eGPU over Thunderbolt3 on its Alienware machines.
Yet, this post is able to do it with an older XPS-15.
Thought that I didn't install the GPU correctly, so remounted the GPU, this time with both powercables attached. Installed Razor Synapse 3. In principal that is only for the lights, but maybe it has a side-effect.
Side-effect or not, now gpuDeviceCount returns 1 in Matlab. Checked the documentation and the default option is auto. If I want to force, I could have added 'ExecutionEnvironment', 'gpu'.
The training-progress screen still says cpu-based. Calling ComputeCapability fails on graphics driver too old (CUDA v7.5 found). Installed Java and run Nvidia SmartScan. Now the Tesla C2050 is found, still the installation of the driver fails. Yet, in my DeviceManager the C2050 is visible, so I did an update from there. Also found some unknown-media's, but the computer restarted before I could update those (only Nvidia audio). Booting is still difficult with the Thunderbolt connected.
The driver of the C2050 is now 21.21.13.6909 from August 2016. Matlab reports that the graphics driver supports CUDA version 8, while it requires 9.1.
Tried to circumvent the driver issue with this advice, but the second option gave the same (Windows has already the best drivers).
The Tesla C2050 is of the Fermi architecture, compute capability 2.0, which means that CUDA until 8.0 is supported.

June 30, 2018

My labtop was put in hibernate by my wife, Matlab asked if it could kill the process. Instead, I activated the pause button (which worked before), but this time no continue option, only Run.
The training has performed 37/120 epochs (3850/12600 iterations) in 3 days time. The result was an accuracy of ~87% and a loss of ~0.2.

June 29, 2018

Ordered a Razer Core V2: looks cool but not too cool, has 500W power supply and extensive interfaces. In addition, a long list of supported GPUs.

June 28, 2018

The third example needed an addin. Addins are not available in a prerelease, but I was redirected to fileexchange. The install package was actually loaded by the addin manager.
The third example hanged on downloading the images. Yet, the url are correct: the dataset is described on Gabriel Brostow from University College London.
After manually creating the directories with the data, the program starts to train (at 09:06). There are 12600 iterations to train, so I am curious how long this takes. According to the ebook, this should take 19h.
During the training, 2 of the cores (4 of the logical cores) are fully used, so the overall CPU usage is 75%
After three hours, only 2.5% of the training is finished, so that would mean 120h (5 days). Time for an eGPU!

June 27, 2018

Downloading Matlab 2018b on my Windows drive on nb-dual. Like to try deep learning code examples for a machine without GPU (but good onboard graphics).
Will see what is possible with external GPU card, where the GPU is provided via NVIDIA grant.
Thunderbold3 is equivalent with USB-C, so this configuration could also work.
The Aorus Gaming Box has Thunderbolt3. Alternative is Akitio Node. Only Thunderbolt3, no GPU installed, but so large (and heavy) that all GPUs would fit. The Asus ROG XG Station 2 also is shipped without GPU, has more interfaces, is lighter, but still huge (and 2x as expensive). My favorite would be Razer Core V2: looks cool but not too cool, has 500W power supply and extensive interfaces.
The blog how to use Thunderbold3 with a Mac with only Thunderbold2, by using a converter and a Mac-script. The script is published on github. An equivalent script exist for Ubuntu.
On egpu.io posts of people getting it working are posted. A system close to mine system is 9360, Ubuntu 17.04, AKiTio Node and a GTX 1080.
Started the first training section of Demo1 of the deep learning code examples at 21:39:04. After 5 minute the training was halfway (104 of the 210 iterations, 15 / 30 epochs, 7 iterations per epoch). As said in the ebook, not much progress is made Actually, MATLAB was training two networks at the same time. Also the one with the correct learning rate was learning (goin up to an accuracy above 90%. After 12 minutes the program started to validate the accuracy on the test-set (and started another training). The last training costs 28 minutes to reach an accuracy of 99%.
The second example had no notebook.
The third example needed an addin. Addins are not available in a prerelease, but I was redirected to fileexchange. The install package was actually loaded by the addin manager.

June 21, 2018

Followed Turtlebot tutorial and performed img = receive(imsub); figure imshow(readImage(img));. Yet, receive never returns.
Restarted Gazebo, received Failed to load plugin /home/arnoud/git/RoboCup2018RVRL_Demo/devel/lib/libgazebo_ros_thermal_camera_pink.so: No so file. Still receive(imsub) doesn't return. The spawn also complains: Could not find parameter robot_description on parameter server.

Cloned RoboCup2018RVRL_Demo.
Couldn't perform catkin_make because catkin was not installed, yet I cannot install it due to dependence on python-catkin-pkg which is installed but seems not be recognized. Broken packages.
Run both the scripts catkin.sh and kinetic.sh in my ~/bin, but still no catkin_make.
Strange, kinetic.sh contains the right source command, but manually performing source /opt/ros/kinetic/setup.bash solves this issue.
catkin_make failed on Add the installation prefix of "hector_pose_estimation" to CMAKE_PREFIX_PATH . This package is available for kinetic.
Put this package in my catkin_ws, but making this packages fails on missing geographic_msgs. That package also exists for kinetic. Making it with command catkin_make --source common_msgs, followed by catkin_make --source hector_localization.
Sourcing ~/catkin_ws/devel/setup.sh made it possible to do rospack find hector_localization, but still catkin_make failed. Made symbolic links to the missing packages.
Could have better installed it with sudo apt-get install ros-kinetic-ros-control. The ros-control couldn't be build. Removed the link in src, performed cleanup, but also had to remove all ros-control directories in ~/catkin_ws/devel/share.
Had to do the same with sudo apt-get install ros-kinetic-geometry2, including removing ~/git/RoboCup2018RVRL_Demo/devel/share/tf2*.
There was an install file on the github (only no instructions to run it). Missing packages at my side: libignition-math3 libsdformat5 pavucontrol plus 57 others.
Performing roslaunch robot_description spawn_multi_robots_pre_indoor1_1.launch didn't work, but roscd robot_description, followed by roslaunch ./spawn_multi_robots_pre_outdoor_1.launch (one 1 less) works. Same for roslaunch setup world_pre_indoor1_1.launch. The command roslaunch launch/world_pre_indoor_1.launch fails on missing RoboCup2018RVRL_Demo/src/setup/worlds/map_pre_indoor_1.world. And that is correct, only two outdoor worlds are available in setup/worlds. roslaunch launch/world_pre_outdoor_1.launch works. Launched four Pioneers in Outdoor1.
Looked at roslaunch kobuki_keyop keyop.launch, which was used in Ros Turtlebot Tutorial. kobuki's keyop.launch has two remaps: and . Our Pioneers, don't have motor_power, the controls are robot0/cmd_vel till /robot2/cmd_vel (no robot3 !). Created a robot0.launch. Could ccontrol the robot (2nd from top). robot1.launch controls the 3rd from the top. The top robot could be controlled with ground.launch. robot2.launch controls the last robot.
Continued with Matlab's Get Started with ROS. The command rosinit works. The command rosnode list gave a much shorter list than rostopic list. For instance, I am missing the cmd_vel and camera_ros. /gazebo
/ground/joint_state_publisher
/ground/robot_state_publisher
/matlab_global_node_79241
/robot0/joint_state_publisher
/robot0/robot_state_publisher
/robot1/joint_state_publisher
/robot1/robot_state_publisher
/robot2/joint_state_publisher
/robot2/robot_state_publisher
/rosout
Yet, rostopic list in Matlab gave the same list.
Followed Turtlebot tutorial and controlled one of the outdoor1 robots with commands robot = rospublisher('/ground/cmd_vel');, velmsg = rosmessage(robot);, velmsg.Linear.X = 0.1;, send(robot,velmsg);.
Continued with the camera: if ismember('/ground/camera_ros/image/compressed', rostopic('list')) imsub = rossubscriber('/ground/camera_ros/image/compressed') end,

June 15, 2018

Installed on Linux gnome-multi-writer.
First tried Open With Disk Image Writer, but that gave gnome-disks crashed with SIGSEGV in udisks_block_get_preferred_device.
gnome-multi-writer has started writing. The warning of erasing the disk comes before you have selected the image. Yet, two sections are OK, the third not (probably just too small).
Trying to install Ubuntu from the 1.5Gb image downloaded from the Ubuntu site. Yet, I tried both option (Trying Ubuntu, installing Ubuntu), but receive a flashing cursor only).

June 14, 2018

The option of my implementation of shortestpath is now in place, the function can now be called with path_index = PG.shortestpath(M(char("36373")),M(char("33976")),'algorithm', 'Dijkstra'). The answer should be the same for Dijstra and Astar (only a longer time to find a solution):
path_index = Columns 1 through 26 2068 1989 1975 62 2206 1422 219 749 486 1177 885 2211 259 238 465 909 1851 248 396 1875 1876 2335 60 1062 2336 2334 Columns 27 through 37 322 1275 1533 720 714 708 709 1049 1043 698 692
For the case "36373" to "33976" this seems to be the case. Now breath-first, then tic-toc measurements.
With breadth-first another solution is found:
path_index = PG.shortestpath(M(char("36373")),M(char("33976")),'algorithm', 'breadth-first') Calculating shortestpath with breadth-first algorithm

path_index =

2068 1989 1975 1857 1855 1853 1850 1845 1843 1838 1836 1831 1827 1825 1821 194 1023 1087 547 701 708 709 1049 1043 698 692
Less steps than Astar and Dijkstra. Should compare the distance. Path is indicated in yellow:
The cost of the yellow path is 260051, while the cyan path from Dijkstra/Astar is 252553.
When I use depth-first, I even get another option, with intermediate costs of 256829: path_index = 2068 1989 1975 1857 1855 1853 1850 1845 1843 1838 1836 1831 1827 1825 1821 1819 369 1342 1898 271 1022 1321 697 698 692
Path of depth-first is indicated in magneta:
Depth-first is not guaranteed to have the optimal solution, but breadth-first should. Should check if the break is not too early (not all neighbours searched)?
Timing the 4 algorithms:
1. Calculating shortestpath with breadth-first algorithm Elapsed time is 0.030769 seconds.
2. Calculating shortestpath with depth-first algorithm. Elapsed time is 0.087983 seconds.
3. Calculating shortestpath with Astar algorithm. Elapsed time is 0.095392 seconds.
4. Calculating shortestpath with Dijkstra algorithm. Elapsed time is 0.202984 seconds.
Seems that I have mixed breadth and depth-first. Breath-first could be faster than Astar and Dijkstra, because it doesn't sort.
Added some code to continue searching for better paths, but none is found (because (vgoal is now in the closedList?).
Tried that, but that gives an endless loop!
Tried to continue with changing the cost. Yet, nowhere is specified how fast an agent can move. Finally found an indication in Yuki Miyamoto's github, Agents move on average 7000, while Civilians move 2000 per turn. So, multiplying the travel cost with a factor 4 seems to be a good strategy.
Received 3 Precision 5820 for the virtual, and 10 3629 precisions.
The Precisions 5820 have a Xeon W-2123 CPU @ 3.60 Ghz with 4 cores. They have a GTX 1080 onboard (and another GPU, with DVI). Has an 236 GB harddrive, so time to download a Linux image.
Ubuntu 16.04.04 image is 1.5 Gb big.
Followed the instructions on linux.datanose.nl and also downloaded USBWriter.

June 13, 2018

Started Matlab in the Joint-Rescue-Forces directory.
cd tutorial/scripts
load ../data/testmap_graph.mat
[PG,M] = graphToGraph(G)
plot(PG,'NodeSize', 3);title('RVC Toolbox PGraph');
path_index = PG.Astar(M(char("279")),M(char("896")))
PG.highlight_path(path_index, 'EdgeThickness', 3, 'EdgeColor', 'cyan')
PG.highlight_node(path_index(end), 'NodeFaceColor', 'red')
PG.highlight_node(path_index(1), 'NodeFaceColor', 'green')
So I reproduced the two paths from April 29: path_index_first = PG.Astar(M(char("976")),M(char("938"))) and path_index_second = PG.Astar(M(char("976")),M(char("905"))). Result:
Wrote a function getTop10connectivity. The node of Kobe with the most connections is:
1. 10 connections: "36373"
2. 9 connections: "33976"
3. 8 connections: "35378"
4. 8 connections: "34146"
Tried to plan a path from #1 to #2-#4 with the code from = find(G.Nodes.ID=="36373");target = [find(G.Nodes.ID=="33976") find(G.Nodes.ID=="35378") find(G.Nodes.ID=="34146"); followed by [TR,D] = shortestpathtree(G,from,target,'OutputForm','cell');.
The path from "36373" to "35378" is clearly the shortest (9 steps, distance 68434). short_path = TR{2};
The IDs of the shortest path could be printed with G.Nodes.ID(short_path) and highlighted with PG.highlight_path(short_path, 'EdgeThickness', 3, 'EdgeColor', 'cyan');PG.highlight_node(short_path_(end), 'NodeFaceColor', 'red'); PG.highlight_node(shortest_path(1), 'NodeFaceColor', 'green');.

June 6, 2018

A pity that we don't use Unreal anymore, with this Science Lab.
Looked at this video tutorial. This shows five steps:
- Download an object from a marketplace.
- Put the object in the origin with SketchUp (Mac), and do some measurements on the dimensions for the sdf of Gazebo).
- Give the object a center of mass in Blender.
- Create a directory in gazebo/models, with as mesh the downloaded file, and the collected measurements into the sdf.
Should find fig 3.14 (Uniform-Cost-Search) from AMAI, because the implementation in the ADF looks to follow those steps.

June 1, 2018

Asked for three servers with GTX 1080 Ti GPU, based on the score according to Tom's hardware.

May 31, 2018

Updating Team Description Paper with the comments of the reviewers and the results of the session in Magdeburg.
Received from Sebastian the results of the "thinning" algorithm:

May 13, 2018

Updated until 2013

May 12, 2018

Published Rules v1.2
Updated 2006
Updated 2007
Updated 2008

May 7, 2018

Francesco sent an article with an after-competition of DARPA challenge.
Another interesting paper is from Leipzig, with has several publications on the Multi-Agent Flood (MAF) algorithm. Seems to be an approach inbetween the agent and the virtual league.

April 30, 2018

Installing on nb-udk the rcrs-server from github. Gustavo has run several training session on this machine.
Cleaned up my code. Nice, but strange enough no output, nor agent movement after the begin of the run. The same code works fine at Gustavo's machine. Other module.cfg.
Added the nodes as vectors (currently BATCH_SIZE = 25). Added tic-toc measurements. For Kobe the numbers are:
nodes added in 1.6564 s
edges added in 30.2186 s
file saved in 0.2883 s
. Precomputing at the map Kobe can be started with the command ./start-precompute.sh -m ../maps/gml/kobe/map/ -c ../maps/gml/kobe/config/. The agents can be started by ./precompute.sh 1 0 1 0 1 0 localhost.
Strange enough, it takes ages to finish after my precompute function is ready.
For BATCH_SIZE = 250:
nodes added in 0.6677 s
For BATCH_SIZE = 2500:
nodes added in 0.42233 s
The original code, without vectors, took:
nodes added in 15.7999 s
The map of Kobe is 593Mb, the TestWord was 3.8Gb.
For the map of Paris, the numbers are (BATCH_SIZE = 2500): nodes added in 0.70695 s
edges added in 51.6891 s
file saved in 0.3587 s
For the TestWorld (BATCH_SIZE=2500):
nodes added in 0.11392 s
edges added in 1.5327 s
file saved in 0.44342 s
Each edge-call takes 10 ms.
Sebastian looks if it is possible to convert the graph to Peter Corke's PGraph, which contains an Astar implementation. No Dijkstra, breadth-first or depth-first, but source-code is available, so that can easily be made. Peter Corke's code can also be converted to C-code, while the MathWorks Graph cannot. Downside of Peter Corke's approach is that nodes and edges have to be added one-by-one.
Peter Corke's AStar code is directly available in matlab-code. An option can be set to change it to Dijkstra algorithm (by setting an additional option in tb_optparse) and putting an if-statement around adding the heuristics in the h_score.
Read this short tutorial, with an explanation of Breadth-first, Dijkstra and A-star. Red Blob uses this pseudo-codes:
% A-star while not frontier.empty(): current = frontier.get() if current == goal: break for next in graph.neighbors(current): new_cost = cost_so_far[current] + graph.cost(current, next) if next not in cost_so_far or new_cost < cost_so_far[next]: cost_so_far[next] = new_cost priority = new_cost + heuristic(goal, next) frontier.put(next, priority) came_from[next] = current
so A-star adds + heuristics() to
% Dijkstra while not frontier.empty(): current = frontier.get() if current == goal: break for next in graph.neighbors(current): new_cost = cost_so_far[current] + graph.cost(current, next) if next not in cost_so_far or new_cost < cost_so_far[next]: cost_so_far[next] = new_cost priority = new_cost frontier.put(next, priority) came_from[next] = current
So Dijkstra adds the priority to breadth-first
%Breadth-first while not frontier.empty(): current = frontier.get() if current == goal: break for next in graph.neighbors(current): if next not in came_from: frontier.put(next) came_from[next] = current
So with first-come, first put strategy. That would mean to take the first from the openSet in line 1314, not the element with the element with the lowest score.
The path to Peter Corke's PGraph.m implementation is ~/Documents/MATLAB/Add-Ons/Toolboxes/Robotics Toolbox for MATLAB/code/common/PGraph.m
Sebastian's graphToPGraph was able convert TestWorld, Kobe and Paris in resp. 0.46s, 12.48s and 21.98s:
The command [PG,M] = graphToPGraph(G) gives also a map between the id and the index. I was able to plan a path with the command path_index = PG.Astar(M(char("902")),M(char("903"))), which gave as answer string(PG.vdata(path_index(2))) the node inbetween "273".
The command path_index = PG.Astar(M(char("279")),M(char("896"))) gave as answer "971" and "970", as tested on April 29.

April 29, 2018

Looked at the path requested by the agents. One is 902 to [903, 904, 910, 974, 911, 912, 336, 913, 914, 915, 916, 917, 918, 919, 920, 921, 922, 923, 924, 349, 925, 926, 927, 928, 929, 932, 933, 297, 314]
[AStarPathPlanning] the shortest path from this collection seems to be:
0: 902
1: 273
2: 903
Another interesting one is 279 to [896, 902, 903, 904, 968, 330, 910, 974, 911, 912, 336, 913, 914, 915, 916, 917, 918, 919, 920, 921, 922, 923, 924, 349, 925, 926, 927, 928, 929, 932, 933, 297, 314]
[AStarPathPlanning] found a path:
0: 279
1: 971
2: 970
3: 896
Most promising seems to be sortestpathtree, which gives for a number of targets to the shortest distance and the path. It is possible to specify the algorithm (Breadth-first or Dijkstra).
d = distances(G) gave a 95x95 matrices with all distances.
Continued with asking the path between 902 and 903. The code is simple:
from = find(G.Nodes.ID=="902"); target = find(G.Nodes.ID=="903"); [TR,D] = shortestpathtree(G,from,target,'OutputForm','cell'); D = 43922 TR{1} = 8 21 10 G.Nodes{21,:} = "273" "29898" "59056" So it found the path 902 -> 273 -> 903.
Now trying 976 to [905, 249, 298, 938, 939, 940]. That means only to change two lines of code:
from = find(G.Nodes.ID=="976") target = [find(G.Nodes.ID=="905") find(G.Nodes.ID=="249") find(G.Nodes.ID=="298") find(G.Nodes.ID=="938") find(G.Nodes.ID=="939") find(G.Nodes.ID=="940")]
The result is
TR = 6×1 cell array {1×12 double} {1×10 double} {1×8 double} {1×9 double} {1×14 double} {1×13 double} D = 159942 176209 166612 158496 181206 175706 [B,I] = sort(D); short_path = TR{I(1)}; G.Nodes.ID(short_path) ans = ans = 9×1 string array "976" "975" "974" "262" "280" "259" "297" "911" "938" That is actually different from the path found by the AStar.java implementation. That path is the second longest path G.Nodes.ID(TR{I(2)}) (length 159942 vs 158496):
ans = 12×1 string array "976" "975" "974" "262" "280" "257" "270" "263" "278" "271" "910" "905"
Implemented this algorithm in Java. Seems that the correct paths are returned, but the PoliceAgent is not moving ?!
Changed that the matlab engine is part of the class, yet the result is that the classic AStar fails on Index-0 in the Node implementation.
Moved the Matlab-code into a Matlab-function, which is called only once. This means that the targets have to be put in a StringArray (StringArrayList failed) and that the path has the be read out as an DoubleArray.
Found that the path was stored in reversed order (from target to from). After that change the PoliceAgent started to move!
Called roslaunch turtlebot_gazebo turtlebot_world.launch world_file:=/opt/ros/kinetic/share/turtlebot_gazebo/worlds/playground.world and now the world has objects. Controlled the robot with roslaunch kobuki_keyop keyop.launch, as specified in ROS Turtlebot Tutorial.
Even activated Rviz for Turtlebot (only had to adjust image_rect_color to image_raw - and activate Image & PointCloud) with roslaunch turtlebot_rviz_launchers view_robot.launch. For all the activation of the scripts source kinetic.sh; source gazebo.sh; source catkin.sh are essential. Result:

April 28, 2018

Sebastian helped with roslaunch (not clear how). Actually, the turtlebot_world is actually in the ros-kinetic-turtlebot-simulation package. Even worse, it is not an actual world, but a enviroment variable or a commandline option. The choice is between empty.world, corridor.world or playground.world. Loading the corridor.world for the first time took ages, playground even longer.
I received the branch of Gustavo. His KMeans part works fine, but my Pathplanning routine is not called (although it is specified in the config). Seems to be overruled in the adf-core, but that was a wrong hypothesis. Gustavo found the error: there was still the .sample. directory before the A*-function in the config-file.
Added a routine to ask the location where the Matlab stored the graph. That was in the location where the program was launched. Should change it to precomp_data.
Made a matlab_hd.sh after checking the available resolutions. 2048x1536 had two big black bands on the side.
Changed that only the FireBrigade made a graph. Still, Matlab goes out of memory very vast. Already over 23 nodes it starts complaining that graph.mat is bigger than 2GB. Yet, no edges were added yet!
Problem is that the graph reserves memory for all members. The agents have 5G ids, so the graph gets huge (and very sparse). Solution is to ask for roads and building only (see SampleKMeans). In addition, the graph is not that interesting, but I found out that is because a grid is made from all non-used IDs. Added the edges, the graph consists of 204 endnotes and 67758 edges, which seems to be a lot. Node 975 has for instance 4 neighbors. When the edges are added, a small structure in the corner became visible.
With the help of Sebastian made a graph with uses a sequention as index, with additional IDs to find the Entities. Also included the X,Y location, so the map can be not only drawn topological but also metrical:

April 27, 2018

With the help of Gustavo, I was able to start three sample agents on the test-map. The trick was to use launch -all, instead of start (which had a bug, becaused it use the pre-parameter as host). The sample agent do not move on the map, because their decision function gets in a loop due to the surroundings.
All documentation should be at rescuesim.robocup.org.
Started with the example to create 'own code': change from Dijkstra path-planning to A*. The next step would be to change this to Matlab code. Many sample implementations are available, native to Matlab is Graph Shortest Path.
Modified the module.cfg to jrf_module.cfg which loads jrf A*. Also created a sample/SampleLoader.java to know the name of the JRF-team. Couldn't compile, because no makefile is provided (and several jars from library/rescue should be loaded).
uva-rescue used ant default to compile. Ant got its instructions from build.xml. Adding the adf-core.jar is simple, but no jar is made from the adf-sample.
Build an adf-sample.jar by the command jar cf adf-sample.jar adf/sample/*/*.class adf/sample/*/*/*.class.
The SampleLoader can than be compiled by javac SampleLoader.java -cp ../../../library/rescue/adf/adf-core.jar:../../../build/adf-sample.jar. Moving this compiled file to ../../../build/jrf/sample made that I could execute my first run (Police started to move to the FireBrigade, but got stuck in the middle). First score 0.0
Continue with Matlab API for java. Run the EngineConsoloDemo.java succesfully.
Looked at the PathPlanning interface (which is loaded from adf-core.jar). The source-code could be extracted with jar -xvf adf-core-sources.jar.
The PathPlanning gets as parameters AgentInfo, WordInfo, ModuleManager and DevelopData. As function you get set a single or a list of Destinations, and get the result (a list of EntityIDs). The work is done in the function calc(). getDistance gives the length of the path.
The current Java A* implementation uses a Map and a Graph. The graph is the variable that should be replaced by its Matlab equivalent. In init, the graph is initiated with a LazyMap (from apache-commons) of neighbours. This graph is only used in the last step of A* (step // 7), where the administration is done (unknown neighbours to the open list, when improvement add neighbour to the hash. The actual path is already filled in step //6 (if the node is the target).
Created a version of AStar.java which loads the Matlab java-engine. It compiles, but it seems that this version is not called (just the adf-version). At least, no system.out from the init().
While waiting on Gustavo's working version, I tried out the Turtlebot example. Downloaded the Support Package for TurtleBot Based Robots, but it is not clear what or where something is installed.
The connecting to an existing gazebo simulator points to the ROS website, but at the end I am redirected to this tutorial. No tutirual for kinetic, but the install works fine for ros-kinetic-turtlebot-apps.
Failed on running / installing roslaunch. That is dependent on python-roslib, which is depended on python-catkin-pkg, yet catkin seems not to be installed (while I started kinetic.sh, gazebo.sh and catkins.sh). Strange.

April 23, 2018

Should try the ROS/Matlab/Simulink/Gazebo tutorial (with Usarsim and/or youBot).
Tried to follow the instructions of Future of Rescue workshop, but this instructions are already outdated for me.
My XPS13-dual has Ubuntu 16.04, so I should install ROS Kinetic, instead of ROS Indigo.
Had some problems with the command rosdep update, but this was because Matlab also had a libexpat.so.1 version. Used the trick of stackoverflow and emptied LD_LIBRARY_PATH.
Used apt-cache search ros-kinetic-gazebo | grep pkgs to see which options were available. According to gazebosim.org I should use gazebo7, so I assume that sudo apt-get install ros-indigo-gazebo-ros-pkgs installs version7 (gazebo8 and gazebo9-ros-pkgs were also available). The command gazebo --version showed version 7.12.0.
Tried to do rosrun gazebo_ros gazebo but had to install rosrun first. Yet, rosrun is missing rospack. Seems to be because I haven't run source ~/catkin_ws/devel/setup.sh. Difficulty to perform catkin_init_workspace. Suggestion is to install catkin, but ros-kinetic-catkin is already installed. Running the ~/bin/kinetic.sh solved this.
Continued with day 1.
Skipped usarsimros instructions and jumped to Shimizu's RoboCupRescuePackage. Installed protobuf-compiler. Tried a make, but get many errors from /usr/include/sdformat-4.4 and /usr/include/gazebo-7/. Strange, because Masaru worked on Gazebo7compatibility.
Could do rosrun gazebo_ros gazebo and gazebo USARGazebo.world.

April 20, 2018

Read a PhD-thesis of Antoine Bautin (in French) about multiple robot mapping. The 3D map the constructed from the rooms of the Carotte competition (2012) looked very impressive.
If I want to learn more, I can better read MinPos : A Novel Frontier Allocation Algorithm for Multi-robot Exploration.

April 18, 2018

The svn-server u013154 was still in the lab. Installed it in C3.157.
Installed on nb-dualboot svn with sudo apt-get install subversion.
Continued with the steps from ubuntu to create a server.
The Users and Groups app didn't exist anymore, so I created a group with sudo addgroup subversion, followed by sudo adduser www-data subversion.
Copied both the svn as svn_repository of RescueAgents from u013154.
Was trying to do a svn relocate, but my version of svn was quite old. First performed in the working directory svn upgrade. Checked the location of the repository with svn info | grep URL. Then performed the relocate with svn relocate file:///home/arnoud/svn_repositories/RescueAgents and did a check with svn log and svn update.
In rescue-0_48 sample code can be found to connect to rescue-server.
In first try to make an agent in original/simple_connect. Make failed on missing librescue.
Corrected a lot of missing c-libraries in librescue, but was able to fix this easily. Could make simple_connect, although the test failed. Changing the order of the libraries (-l rescue after test.o) solved that. Commited revision 31.
Failed to build the firesimulator due to unmappable characters in UTF8. Solved this with an hint from stackoverflow: find -name '*.java' -exec recode Latin-1..UTF-8 {} \;
Sebastian recommendents an Euclidian algorithm, an unclustered point cloud clustering algorithm in the Computer Vision Toolbox.
Today's webinar of the ConstructSim shows nicely the essence of multi-robot control in ROS: transform the laser-readings from robot to the frame of another robot. Unfortunatelly, again my project doesn't open.
Till the end of the month we can propose an online ros-tutorial at online ros-developers conference. Maybe I should use how to make a ROS-tutorial.

April 17, 2018

Using vectorizer to create svg.

April 12, 2018

Added transparency to the images for the team-shirts with an online-image-editor.

April 4, 2018

Should try RoboCup@Home Learn workshop.

April 2, 2018

Should try an URDF from a Nao, Pepper or youBot. The URDF of the nao is available on ROS wiki. Descriptions of V32, V33, V40 and V50 are available. Examples of other robots are also available.
Also nice to port the ViSP implementation from Romeo to Pepper.

March 29, 2018

Added Matlab to my path, but ssh has already some conflicts with the crypto libraries, so could better put the LD_LIBRARY in a script.
Matlab doesn't support 4K displays yet, so used the script suggested in the forum.
Moved the LD_LIBRARY to this script.

March 28, 2018

Installed on nb-dual Matlab R2018a on /usr/local/MATLAB/R2018a/. RoboCup license is with all toolboxes, including:
- Automated Driving System Toolbox 1.2
- Computer Vision System Toolbox 8.1
- GPU Coder 1.1
- Image Processing Toolbox 10.2
- Mapping Toolbox 4.6
- Parallel Computing Toolbox 6.12
- Robotics System Toolbox 2.0
- Statistics and Machine Learning Toolbox 11.3
Also installed the MATLAB drive to synchronize Matlab files between my machines.
Here are additional resources for my team:

March 27, 2018

Requested a Matlab license for my dual-boot computer from student competition page.
Following the instructions of Agent competition tutorial.
The first instruction sudo apt-get install openjdk-7-jdk ant rng-tools already fails, could try to follow this patch. Yet, the github readme indicates jdk8.
So, instead, I used the command sudo apt-get install openjdk-8-jdk ant rng-tools.
Also the readme contains errors: had to use git clone git@github.com:roborescue/rcrs-server.git instead of git clone git@github.com:roborescue/rcrs-server.git.
Enough warnings from checkstyle and findbugs, but build is claimed to be successful.
Both the test world and sampleagent script start, but I don't see any time running.
The readme says ./start.sh ../maps/gml/test while the manual says ./start.sh -m ../maps/gml/berlin.
The readme says ./sampleagent.sh, while the manual says sampleagents.sh.
Sampleagents.sh doesn't exist, berlin does. Launching the sampleagent fails:
INFO launcher : Processing ../jars/misc.jar DEBUG launcher : Inspecting manifest... DEBUG launcher : Looking for likely class names... INFO launcher : Launching many instances of component 'sample.SampleFireBrigade'... INFO launcher : Launching instance 1... INFO launcher : failed: No more agents INFO launcher : Launching many instances of component 'sample.SampleAmbulanceTeam'... INFO launcher : Launching instance 1... INFO launcher : failed: No more agents INFO launcher : Launching many instances of component 'sample.SamplePoliceForce'... INFO launcher : Launching instance 1... INFO launcher : failed: No more agents INFO launcher : Launching many instances of component 'sample.SampleCentre'... INFO launcher : Launching instance 1... INFO launcher : failed: No more agents.
Continued with Agent Development manual.
Instead of the downloading the zip-file, I downloaded the latest version with git clone https://github.com/RCRS-ADF/RCRS-ADF.git. Yet, the result is a directory with two empty directories.
Downloaded wget https://github.com/RCRS-ADF/RCRS-ADF/releases/download/release/adf-sample.tar.gz, followed by java -jar ./library/rescue/adf/adf-core.jar -compile. 0 warnings.
Start script works, but seems that 0 agents are launched:
[ RCRS ADF Version 2.2.0 (build 20170726T064111Z) ] [START ] Connect to server (host:localhost, port:7000) [FINISH] Connect FireStation (success:0) [FINISH] Connect PoliceOffice (success:0) [FINISH] Connect AmbulanceTeam (success:0) [FINISH] Connect PoliceForce (success:0) [FINISH] Connect FireBrigade (success:0) [FINISH] Connect AmbulanceCentre (success:0) [FINISH] Done connecting to server (0 agent)
Read 2014 Labbook and saw that I run Mircea's code on Feb 14 on svn/RescueAgents. Latest update is from August 5.
The repository was located on u013154.science.uva.nl/RescueAgents.

March 15, 2018

Read the Tutorial Robot Manipulator Capability in MATLAB by Peter Corke. Online is more material available.

March 6, 2018

The World Robot Challenge also has a Disaster Challenge. This challenge is based on Choreonoid (http://choreonoid.org/en), which can provide the simulation of a high-accuracy virtual disaster environment. Choreonoid is based on Ubuntu 16.04. It contains two robot models, the spider-like WAREC-1 and the double-arm Osaka robot. It also contains several disaster environments and tasks.
The Choreonoid environment is available on github.

March 5, 2018

Logged in at rescuesim.robocup.org and uploaded the new picture I received from Masaru.

February 15, 2018

Read Gazebo blog about the update to Gazebo 9.0.0. This seems a major update, with most of the middleware move to the Ignition libraries.
There is now also Gazebo support for Maritime RobotX competition.

February 13, 2018

Lat91 is just J.-C. Latombe. Robot Motion Planning. Kluwer Academic Publishers, Boston, 1991, the textbook next on my list to check.
The vertical decomposition is based on Chazelle's Approximation and decomposition of shapes, but that is only available as Google book (with missing pages.
An alternative coulde be Minimum Decompositions of Polygonal Objects.

February 12, 2018

Found this technical report about Exact Cell Decomposition, which refers to Lat91 as source of the Exact Cell Decompositiona as described on page 377 of Siegwarts' book.

February 8, 2018

Looked at the 3min summary video of the Matlab Machine Learning Toolbox.

February 7, 2018

Looked into the several clustering algorithms available in Matlab. Hierarchical clustering is already part of Matlab before 2006, and uses a divise algorithm, probably DIANA (DIvisive ANAlysis Clustering). For details, they point to the textbook. Looked another texbook, which clearly explains the clustering techniques, but gives no hint about the history of hierarchical clustering.

February 6, 2018

Checked the deadlines of the Lorentz Center: 30 January, 30 May and 30 September. Workshops taking place sooner than 16 months after the deadline will be evaluated. Thus, the upcoming deadline, 30 May 2018, is for workshops with starting date before 1 September 2019.
Meeting with Sebastian and Luis. Moved the visit to 19 to 23 March.
Luis run the same map 10x with another seed. He added some extra information (initial distance, remaining HP in refuge).
Try to classify HP classes
Try to predict end distances
Try to predict final HP
The building.csv only contains information known to the agents, the damage is only seen when the approach the building. Yet, it doesn't change during the run.
Ententies are both roads and buildings, the connections are not included. The damage of the roads is not included. In the pathplanning, the links are removed (not increasing the weight).