Robotic Arm

Dear all,

Me and a buddy of mine are thinking about trying to make a robotic arm. Does anybody have any useful hints for getting started. I am an electrical engineer with a little software experience, and he is an IST major with fairly extensive experience programming in many languages.

Any help would be greatly appreciated.

As in something you strap onto a person's shoulder or something you lift a car body with?

In theory something you lift a car body with, but it won't lift anything that heavy. we want to first start by making it move, wave, gestures. That will take us a while. Once we get that far, we'll see where we want to go from there.

Big project

Nick and Gigs said:

Dear all,

Me and a buddy of mine are thinking about trying to make a robotic arm. Does anybody have any useful hints for getting started. I am an electrical engineer with a little software experience, and he is an IST major with fairly extensive experience programming in many languages.

Any help would be greatly appreciated.

Click to expand...

You need to learn how to rotate vectors in 3D and write target or motion profile generators. I bought some robotics books.

I'd go like this:

Pick a set of parameters required to match (accel, load capacity, axis count)

Compare this to budget / resources.

Develop your kinematics model and apply your loads. Plot some of your common motion profiles and study the loading characteristics.

From this, size your motors/gear ratios.

Develop your inverse kinematics model/algorithm. Depending on your loop control mechanisms I'd go w/ Genetic algorithm Neural net solving for path planning and simple PID Control for your positioning, I find this to be the most robust.

Then go build it!

You may want to check out Servo Magazine and their web site and forums:
http://www.servomagazine.com/

There is some mention of coordinate transformations in this thread:
http://www.plctalk.net/qanda/showthread.php?t=26616&highlight=Orbiter
But all the links to the old patchn site seem broken. I'll have to dig up the files on an old hard drive.

Norm, I still have your orbiter animations

Sometime I need to learn how to do those animations.

Nick and Gigs, how many degrees of freedom do you plan to have? The degree of difficulty goes up exponentially with the number of degrees of freedom.

We get involved with robotic applications. We have a few that handle nuclear material. Even something that sounds simple like racking drill pipe requires 3 degrees of freedom.

Honga, how can neural nets help with trajectory control? Most robot are taught by having a teach mode where the actuators are jogged to a position and the actuator positions are save. There is the kinematics aren't too difficult in this case but the real trick is to go from point to point at a constant velocity and smoothy ramping through the points. This math gets to be pretty involved.

This is purely from a home build type of view, not using any commercial software.

I use neural nets to control the dynamics of the joints based on giving the robot a point in space to go to.

If you give your robot time to self train with some sort of cost based decision making it can become quite robust. So programming motion paths is irrelevant.

I've managed to get a robot with a 500mm radius accurate to within a 5-6mm^3

?????

Honga said:

I use neural nets to control the dynamics of the joints based on giving the robot a point in space to go to.

Click to expand...

Did you use commercial software?

If you give your robot time to self train with some sort of cost based decision making it can become quite robust. So programming motion paths is irrelevant.

Click to expand...

It doesn't sound like you programed this yourself. Optimizing neural nets require an optimizing/minimizing routine similar to what is used to auto tuning.
So what is Nick and Gigs to do?

I've managed to get a robot with a 500mm radius accurate to within a 5-6mm^3

Click to expand...

Hopefully that is all mechanical error. You should be able to do much better with the pure math.

It should be possible to compute an exact motion path that moves through a series of teach points. It is also possible to figure out the the motion profile so that the the absolute speed, V=sqrt(Vx^2+Vy^2+Vz^2) is constant. Then the trick is to compute the actuator positions that result in the X,Y,Z position as a function of time. This is an iterative process involving Jacobians. Since one position doesn't vary much from another it doesn't take many iterations to get to the next position. These techniques one can find in a book on robotics. Generating a motion path requires only math and no neural nets.

What do you use the neural nets for?

I found all my old stuff on the Orbiter.

The 2 main animations of the motion are here:
http://www.flickr.com/photos/njdjr/sets/72157624800296311/

Kinematics files are attached. There is a brief description of the way the coordinates are transformed from one frame to the other.

Neural nets are mainly a wank. But looking at the free lunch paradigm I should be able to accomplish different things than your normal robot arm.

I'm still just a Uni student, so lots of spare time to think/code.