[Home]Ikfast

LinuxCNCKnowledgeBase | RecentChanges | PageIndex | Preferences | LinuxCNC.org

This page describes how to use the ikfast inverse kinematics module from openrave system.

From the openrave web site, "OpenRAVE? provides an environment for testing, developing, and deploying motion planning algorithms in real-world robotics applications." For linuxcnc, we are using the ikfast module to create the IK component, and the qtcoin display application to show the robot model.

Source

A modified version of the openrave source is available here, https://github.com/ftkalcevic/openrave, in the "linuxcnc_gen_c" branch. See the openrave web site for details of the build dependencies, and how to build and install the package. (http://www.openrave.org/docs/latest_stable/coreapihtml/installation_linux.html)

The modified version adds a c code generator that works with linuxcnc. Use the --outputlang=c to create the c code.

The ikfast updates to linuxcnc can be found here, https://github.com/ftkalcevic/LinuxCNC, in the "ikfast" branch. There is a new src/emc/kinematics/ikfastkins.c module that uses the ikfast generated code. There is also some example configurations in configs/ikfast-demo directory.

Building a robot model

openrave supports multiple file formats to describe the robot, http://www.openrave.org/docs/latest_stable/robots_overview/. Openrave xml is a simple format that allows models to be generated quickly. The COLLADA format allows you to generate more realistic looking robots.

upload:ikfast1.png
Simple robot defined using openrave xml

upload:ikfast2.png
The same simple robot, but defined using collada file

The robot model is used in all openrave computations, including the generation of the ikfast source. Therefore it is important that it is created accurately.

The creator of openrave recommends opening an interactive openrave session and continually reloading the model...

$> openrave.py -i
In [1]: env.Reset(); env.Load('simple.robot.detailed.xml')

Generating and testing the Inverse Kinematics

The recommended way of generating the ik code is to use the openrave.py program, but advanced users can use ikfast.py directly. It is best to generate the cpp component first and use that for automatic testing. Automatic testing doesn't work with .c output.

For example, to generate the 3d translation ik...

$> openrave.py --database inversekinematics --robot=../simple.robot.xml --iktyp
e=Translation3D --freejoint=Wrist --freejoint=WristR --numiktests=1000

getIndicesFromJointNames [3, 4] ['Wrist', 'WristR']
openravepy.databases.inversekinematics: generate, Generating inverse kinematics 
for manip arm: Translation3D [0, 1, 2] (this might take up to 10 min)
openravepy.databases.inversekinematics: generate, creating ik file /home/frankt/
.openrave/kinematics.587828202130adcfd9bfdcd543fb17e2/ikfast62.Translation3D.0_1
_2_f3_4.cpp

... Lots more output ....

openravepy.databases.inversekinematics: save, inversekinematics generation is do
ne, compiled shared object: /home/frankt/.openrave/kinematics.587828202130adcfd9
bfdcd543fb17e2/ikfast62.Translation3D.i686.0_1_2_f3_4.so
openravepy.databases.inversekinematics: RunFromParser, testing the success rate 
of robot ../simple.robot.xml 
[ikfastmodule.cpp:946] FindIKSolution: No ik solution found, i = 0
Joint Val: 0 0 0 0 0 
Transform: 855638019 0 0 0.4799 
raw ik command: 0 0 0 0 0 0 0 0 0 0 0 0.4799 0.2606354555368424 0.40057338399291
04 

[ikfastmodule.cpp:1001] FindIKSolutions: No ik solution found for, i = 0
Joint Val: 0 0 0 0 0 
Transform: 855638019 0 0 0.4799 

openravepy.databases.inversekinematics: testik, success rate: 0.999000, wrong so
lutions: 0.000000, no solutions: 0.001000, missing solution: 0.000000

Specifying --numiktests runs a series of tests with the ik module to confirm it works correctly. I suspect the output shown above is a bug, as the first test coordinate looks dodgey. Running the tests is important because many ik solutions don't provide a very high success rate. If this happens you'll need to look at changing the iktype, or possibly selecting different free joints.

For more details on how to use openrave, visit the openrave website, www.openrave.org

openrave.py can also do performance testing of the ik solution. Specify the --perftiming= option.

$> openrave.py --database inversekinematics --robot=../simple.robot.xml --ikty
pe=Translation3D --freejoint=Wrist --freejoint=WristR --perftiming=1000

... Lots of output ...

openravepy.databases.inversekinematics: RunFromParser, running time mean: 0.000
005s, median: 0.000005s, min: 0.000005s, max: 0.000015s

Showing a mean of about 5us to calculate the ik. This is good. It means ikfast produced a fast closed form solution.

Changing iktype to TranslationZAxisAngle4D? gives....

$> openrave.py --database inversekinematics --robot=../simple.robot.xml --ikty
pe=TranslationZAxisAngle4D

... Lots of output...

openravepy.databases.inversekinematics: RunFromParser, running time mean: 0.0027
76s, median: 0.002728s, min: 0.001594s, max: 0.005023s

These are terrible times - a mean of 2.7ms and max of 5ms means we can't run with a 1kHz servo loop. Slowing the servo loop down to 200Hz will produce poor motion.

The slow times are because ikfast had to use a numerical inverse kinematics solver. Numerical solvers are a slow iterative process. This limitation is because of the geometry of the robot and the selection of joints.

To generate the ikfast code for linuxcnc, just add the --outputlang=c option, eg...

$> openrave.py --database inversekinematics --robot=../simple.robot.xml --iktyp
e=Translation3D --freejoint=Wrist --freejoint=WristR --outputlang=c 

getIndicesFromJointNames [3, 4] ['Wrist', 'WristR']
openravepy.databases.inversekinematics: generate, Generating inverse kinematics 
for manip arm: Translation3D [0, 1, 2] (this might take up to 10 min)
openravepy.databases.inversekinematics: generate, creating ik file /home/frankt/
.openrave/kinematics.587828202130adcfd9bfdcd543fb17e2/ikfast62.Translation3D.0_1
_2_f3_4.c

... Lots more output ...

openravepy.ikfast: writeIkSolver, generating c code...
openravepy.databases.inversekinematics: generate, compiling ik file to /home/fra
nkt/.openrave/kinematics.587828202130adcfd9bfdcd543fb17e2/ikfast62.Translation3D
.i686.0_1_2_f3_4.so
openravepy.databases.inversekinematics: generate, cannot continue further if out
putlang c is not cpp
openravepy.databases.inversekinematics: save, inversekinematics generation is do
ne, compiled shared object: /home/frankt/.openrave/kinematics.587828202130adcfd9
bfdcd543fb17e2/ikfast62.Translation3D.i686.0_1_2_f3_4.so

Note the name of the output file in the program logs. The directory name and file name are dependant on the robot model. In this case it is...

/home/frankt/.openrave/kinematics.587828202130adcfd9bfdcd543fb17e2/ikfast62.Translation3D.0_1_2_f3_4.c

To use this file in linuxcnc, copy the file to the linuxcnc source tree...

src/emc/kinematics

and rename it from .c to .inc. Also copy the file ikfastkins_c.h.

Edit ikfastkins.c in this directory to include the .inc file. There should already be a line like...

#include "ikfast62.Translation3D.0_1_2_f3_4.inc"

Just change that one.

Building

Closed Form Solution

If the ikfast code is closed form solution, just do a normal build. (make from the src directory). Be aware however, the generated ikfast solution can be megabytes in size and can 1) take a long time to generate, and 2) requires large amounts of RAM to compile (>2GB).

Numerical Solution

If the ikfast code is a numerical solution, the code needs to compile and link with the lapack linear algebra libraries. Because of the slow speed of the numerical solution, this really only works in simulation, but instructions for a realtime build are also provided.

For a simulator build, edit the main make file (src/Makefile?) and uncomment the line...

#IKFASTLIBS := -llapack

A realtime build is a little more tricky; the kinematics object is a kernel object which has tight restrictions: c only (hence the new ikfast c generator) and no external libraries.

To work around this, there is a patch file, ikfast_fortran_kernel_header.patch, in src/emc/kinematics that has to be applied to the kernel headers to change the kernel makefiles to support compiling fortran. Note - this patch file has only been tested against the linux-headers-2.6.32-122-rtai headers.

Then edit the src/Makefile? and uncomment the very long line that looks like ...

#ikfastkins-objs += emc/kinematics/lapack/lapack/blas/daxpy.o emc/kinematics/lap
ack/lapack/blas/dcopy.o emc/kinematics/lapack/lapack/blas/ddot.o emc/kinematics/
lapack/lapack/blas/dgemv.o emc/kinematics/lapack/lapack/blas/dger.o emc/kinemati
cs/lapack/lapack/blas/dnrm2.o emc/kinematics/lapack/lapack/blas/drot.o emc/kinem
atics/lapack/lapack/blas/dscal.o emc/kinematics/lapack/lapack/blas/dswap.o emc/k
...

The kernel build will now compile the required lapack functions and link them directly with the ikfastkins component.

ikfastkins - the Kinematics Module

The ikfastkins module provides a simple interface to the ikfast ComputeFk?() and ComputeIk?() functions.

Any free joints that are determined by ikfast are mapped to free axis (ABCUVW) so these joints can be moved when in world coordinate mode.

A call to ComputeIk?() will usually return more than one solution. ikfastkins will compare the current joint positions to each solution and pick the closest. This is a simple method that can cause joints to flip if the wrong solution is chosen.

If no ik solution is found, ikfastkins returns GO_RESULT_ERROR, which is ignored by motmod - the robot stops moving, but motmod doesn't, until either the cartesian movement finds a new ik solution, or the robot backs up out of the "deadzone".

For a fully defined 6DoF? robot, there is no interface to define the end effector orientation. Currently that has to be defined when in joint mode, then the orientation will be maintained when switching to world mode.

Display

In the config/ikfast-demo directory is a simple python hal script to display robot model, openravedisplay. The script uses the openrave libraries so they must be installed.

The program will automatically create hal pins for each joint (openravedisplay.joint%d). These can be connected to the hal joint positions.

Eg...

loadusr -W ./openravedisplay --robot=wam.robot.env.xml --refreshrate=15 --showtrail

net j0 axis.0.joint-pos-fb openravedisplay.joint0
net j1 axis.1.joint-pos-fb openravedisplay.joint1
net j2 axis.2.joint-pos-fb openravedisplay.joint2
net j3 axis.3.joint-pos-fb openravedisplay.joint3
net j4 axis.4.joint-pos-fb openravedisplay.joint4
net j5 axis.5.joint-pos-fb openravedisplay.joint5
net j6 axis.6.joint-pos-fb openravedisplay.joint6

Samples

There are some samples in the config/ikfast-demo directory.

Running robotarm.simple will show the robot above. Edit simple.robot.env.xml to change the model. There are 2 ikfast files in src/emc/kinematics that work with this robot...

ikfast62.Translation3D.0_1_2_f3_4.inc - simple 3d translation
ikfast62.TranslationZAxisAngle4D?.0_1_2_3_f4.inc - slow numerical 4d solution

ikfastkins.c must be edited and the #include line changed.

There is also an example of the Barret Wam 7DoF? robot. This robot is included in the openrave installation. The robot can be run using the robotarm.wam script. Remember to edit ikfastkins.c to include the correct ik module (ikfast62.Transform6D.0_1_3_4_5_6_f2.inc)

upload:ikfast3.png

Barret Wam 7DoF? robot

ToDo?


LinuxCNCKnowledgeBase | RecentChanges | PageIndex | Preferences | LinuxCNC.org
This page is read-only. Follow the BasicSteps to edit pages. | View other revisions
Last edited December 12, 2013 9:37 pm by Ftkalcevic (diff)
Search:
Published under a Creative Commons License