6/11/07 – Day 6
Today we used Ghost for Linux to obtain image of Mobile 33. We are able to setup an ftp server and upload the image on there.
It has become clear that there is some sort of internal conflict with USB devices that is preventing all of our peripherals from operating simultaneously. I noticed the package udev fails to launch on boot. The daemon can still be started afterwards however, using the command udevd.
Studying dmesg output reveals that the USB hub we are using registers devices as low speed USB. There are only two USB ports on the labtop, and we have three devices to attach: camera, motor, and IR sensors. Our working configuration has the IR sensors and motor plugged into the hub (low speed USB) and the camera plugged directly into the labtop.
I have been able to successfully run run_client_gui a few times now, with all three peripherals functioning. I believe udev was a large part of the problem, but it is also important that the devices are plugged in to the ports as described above.
I've now used the navtool application successfully several times, and believe the trick is all in the order of testing peripherals. As far as I can tell, if this sequence is followed, the navigation tool will work:
- Plug in motor and IR sensors to hub
- Plug in hub to side USB port
- Plug in Camera to back USB port
- Run test_battery and test_range_sensors
- Run test_camera
- Run run_slam_explore
- Simultaneously, run run_client_gui
At the present time, Mobile 33 works well when these steps are followed. Mobile 37 is temporarily out of commission, but we hope to have a clone of Mobile 33's drive installed early tomorrow. The disk image will be created overnight.
One of the next things I would like to do is calibrate the camera and odometry systems. The odometry system demands a joystick, which we do not have yet, but Ivan has given the go ahead to order one. There are two calibration tools for the camera, an intrinsic and an extrinsic. The intrinsic calibration deals with parameters of the camera itself, whereas the extrinsic calibration involves the camera's positioning on the robot.
I experimented with calibrate_camera_intrinsic today. The program involves taping a checkerboard printout to a flat surface (one of the pillars in Orbit) and positioning the camera in such a way that the software can recognize the squares and corners. Several pictures from different angles need to be taken to calibrate the camera well. We experienced much difficulty getting the software to recognize the shapes. This seems to be at least partially due to the room's lighting. After playing with different distances between the camera and the image, and different ambient lighting scenarios, I was able to take one successful calibration image. I hope to take several more tomorrow.
I'd like to set some goals for this week. First off, I would like to complete calibration for both robots and have them wandering around successfully using navtool. Additionally, I would like for us to write some of our own code, either in c++ or python, that uses the obstacle avoidance feature successfully.
6/12/07 - Day 7
Image ghosting was successful. Not only did we make a duplicate of Mobile 33 onto Mobile 37, we also made a third copy on our recently acquired Mobile 40. For now I will refer to each robot by the fictional robot on its KDE wallpaper, so Mobile 33 = Terminator, Mobile 37 = Bender, and Mobile 40 = Artoo. Maybe we'll come up for better names for them soon.
We had some success calibrating the Terminator's camera with calibrate_camera_intrinsic. The best place we found to do it was the lobby right next to the yellow doors of Winlab, where the light is dim and non-fluorescent. We intend to complete the now mundane task of calibrating the remaining two cameras soon.
We accidentally ripped one of the chords that connects the battery to the motors today. This was an unfortunate setback, and it took up some time to solder new connectors together.
I played around with the teleoperation applet of navtool for a little while. The premise of this software is to allow the user to click on the floor somewhere ahead of the robot in a panel representing the camera's point of view and have the robot successfully navigate to the point clicked. The applet has an odd behavior, in that the camera picture does not show appear in the panel where it is supposed to. It does, however, appear in the adjacent panel labeled "camera local." The applet still tries to work, and if I click somewhere in the blank panel the robot takes off, attempting (presumably) to move to the point where I would have clicked had the picture shown up. I hope to debug this error soon and have the teloperaterion applet functioning properly.
I installed Sun's Java SDK on Bender. This bypasses the java errors we were having before, and I can now run the Behavior Composer using a modified script I called composer.sh.2. The functionality of this program is still a little elusive but I'm sure it will be useful once we get deeper into programming.
I feel like today we are close to completing the Getting Started Guide portion of the documentation. We have yet to calibrate all three robots and still need to figure out teleoperation. In addition, we need to manually edit (for all three robots) the resource-config.xml file which keeps track of the spacial positioning of the cameras, sensors, and wheels of each robot (All three are a bit differently shaped.)
6/13/07 - Day 8
We completed intrinsic camera calibration for all three robots and extrinsic calibration for two of them (the terminator's camera is not mounted properly and we cannot perform calibration until we have some glue to fix it.)
A potentially serious problem we've realized is that the navigation tool is not remembering any landmarks. This is a big problem because landmarks are the way the robots remember obstacles they've encountered in a room and build a map of their surroundings. Originally it was thought that lack of calibration was the problem, but having calibrated the robots we find that at least Bender still does not catalog any landmarks in Navtool. I now suspect the culprit may be a broken communication link between the vSlam applet and the camera. An error in the logs warns us to check permissions on /dev/video0, which are set to 666 as I believe they should be.
Today, magically, the teleop applet was working with camera support on Bender. When we clicked a destination for the robot, it moved in generally the right direction, but passed the target and spent a long time colliding into the wall until we switched off the battery. This is disappointing, especially considering that the camera is supposed to be calibrated.
This afternoon I figured out the landmarks problem. The very simple solution: run run_slam_explore before ever running run_client_gui. The theory: When we first started playing with navtool, we realized that if we started the server first, we did not see any video on the side panel labeled "camera local." We took this lack of video to mean that things were not working properly, and soon discovered that if the client was launched first and the server second, the video appeared. What we didn't understand was that whichever program is launched first hijacks the camera and prevents the other from accessing it. Therefore; when we launch the client first, we see the video, but the slam server cannot take pictures and therefore cannot create landmarks. However, if the server is launched first, we cannot see the video but the slam server can take pictures and can and does create landmarks. Now, when I run the client, if I see the video right away I know something is wrong. (Often times I forgot to turn on the battery).
I spent part of the afternoon reassembling the chassis for Artoo. The previous design was pretty abysmal; there was no place for the motor battery and the labtop couldn't even sit on it unless it was partially closed. I reworked it to solve those two problems, but I fear that the motors might be installed backwards as well. I'm not sure if this is the case, but we'll check it out early tomorrow and resolve the issue if it is one.
6/14/07 - Day 9
As it stands, only Bender is fully operational. Artoo's chassis is looking good, but for some reason the robot still refuses to make landmarks. When on explore mode in Navtool, it moves in a peculiar pattern that made us wonder if there was something wrong with the left motor so as to make it sluggish. The robot tends to travel in a curved path. However, using a python script (triangles.py) we found that the robot is quite capable of moving in a straight line. I suspect that the lack of landmarks is a result of this erratic movement rather than any disconnection between the software and the camera; I remember reading somewhere in the guide that it is difficult for the robot to create landmarks while moving along curved paths. (This is by no means certain)
The thing keeping the terminator from working for now is that the camera is unmounted. We need some epoxy to mount the camera and then we will test it's navigation system.
We learned how to calibrate the IR sensors today. The task involves measuring the position of the sensors with respect to the coordinate system of the robot, which is defined as follows: The origin is located on the floor directly between the two back wheels and directly below their common axis. Positive X is the forward direction of the robot, positive Y is leftwards, and positive Z is up. The parameters Roll, pitch, and yaw are angles defined as rotations about the axes X, Y, and Z respectively. To calibrate the IR sensors, all we need to do is measure these 6 quantities (position and rotation) and declare them in the file resources-config.xml.
We turned our full attention to the Tutorials Guide for the first time today. The first tutorial, 01-simple, is a code that makes the robot move a given distance forward, take a picture, and display the picture momentarily. We got the code to make the robot move and snap a picture, but code creating the GUI picture display refused to compile. The compiler complains about the type !IImageWindow being undefined, although the !API indicates that it is a class in the Evolution namespace. We will investigate this more tomorrow.
It took quite a while before I felt like I understood this first bit of code. The programming is very high level, and many objects need to be created for apparently elusive reasons. For example, one method takes two arguments. Before these arguments can be passed to the method, they must be stored in a TaskArgs object. To complicate things further, one of the arguments is a DoubleArray. A DoubleArray is an array of doubles but stored in an object defined by Evolution. Rather than simply creating an array of doubles, we need to create an array of doubles and pass it as an argument to a new DoubleArray object. This instance of DoubleArray is then passed to TaskArgs which in turn gets passed to a new TaskContextPtr, which is a sort of smart pointer containing the data needed to execute a task. To execute a method, we must pass the result of a TaskContextPtr's get() method to a task functor's run() method. The whole process seems a bit convoluted, but we're hopeful that we'll be able to accept it and use it.
I read ahead a bit in the tutorials guide. Among the many features Evolution provides is an object which handles the processing of multiple tasks simultaneously. This will certainly come in handy when we write our own code.
We've gotten a few suggestions for code we should write, and I think we should get started on it soon. First off, I would very much like to see obstacle avoidance working with our code, as I feel it is in many ways fundamental if the robots are ever to be useful in Winlab. Once this is figured out, (on Ivan's suggestion), we should write some code that makes the robots move about basically randomly (except for obstacle avoidance). Additionally, I would like to make some robust and easy to use MoveToCoordinate type commands which through SSH or whatever other means will cause the robots to create design a smart path and travel to a desired location. With these goals accomplished reliably, we'll be able to start using the robots to study wireless communication.
6/15/07 - Day 10
It's been an exciting day. After a small amount of work, we've confirmed that all three robots can happily wander around Orbit, making landmarks.
Something we realized today was that the sensors on Bender's chassis were mixed up. The East sensor was on the left side and vice versa. We corrected this issue by modifying the sensor locations in resource-config.xml because it would have taken quite a bit of disassembly to physically switch the two sensors.
We're having bad luck compiling our own c++ code. The sample code shipped with ERSP comes with Makefiles which reduce the job of compiling to issuing the command make in the correct directory. We have not been able to successully generate our own make file to compile our code, and are finding that running gcc or g++ directly does not work at all. Many flags and parameters are needed by the compiler. I'm trying to learn how to use the commands autoscan and automake to generate make files automatically, but the going is slow and many errors are appearing along the way.
In the mean time, Vamsi has been playing with python scripting with some great success. He discovered a command which causes the robot to wander around with obstacle avoidance enabled. It seems python has access to much of the ERSP and may actually be enough for what we need from the robots. However, it is not nearly as well documented and might be much more difficult for us to learn.