Sunday, August 26, 2012

VGR movies

If you enjoy watching robots do the random bin picking thing, Danish vision and robotics specialist Scape Technologies has a treat for you. A host of movies, (also on YouTube,) show their single camera vision system, (they snap two images from different poses and do some stereo work to calculate a 3D map,) combined with some clever gripper designs. It all looks very impressive, so allow me to pour a little cold water on their parade.

First off, none of the movies show the bins being completely emptied. In fact they only seem to pick off the upper layers of pretty full bins. I’d be more impressed if they demonstrated picking the last few pieces from the bottom.

Second, it all looks rather slow. Sure, a single camera approach cuts the investment slightly, but at the expense of cycle time. I’d want to just strap on a second camera and save the two seconds that’s spent reorienting and acquiring.

Now in fairness, what they’re doing looks good. I’m just not sure it’s any better than what I’ve seen from other companies. And I think everyone struggles to completely empty the bin. In other words, I don’t believe bin picking is truly ready for prime-time.


Anonymous said...

It's possible to use off-gripper cameras (stereo or laser scanner). There's also option to handle picking with single camera without moving the gripper.

All tasks cannot run at "full speed", the following machines can limit seriously the through put. Jus think how long milling or turning takes time.

System can handle the last pieces (if physically possible), I'll ask if they have video about that..

Anonymous said...

These I could found:

Video of emptying bin (in lab, 2009):

Video of getting layer almost empty, the one in the corner is hard to detect of pick, don't know which is case here (real life, 2008):

Getting bin empty isn't critical in this application, since pallet change is manual.

René Dencker Eriksen said...


This is Rene Dencker Eriksen, CTO of Scape Technologies. I'm happy to see that there is some discussion going on about bin-picking. Let me address the questions raised in the original post by B Grey:

* It is impossible to promise that we empty a bin completely. There is always a risk that a part is "clamped" into the corner of the bin and that we can't get the gripper into the corner. Another situation is that a relatively flat part is leaning on the side of the bin wall. In this case we get no sensor feedback because the surface of the part is almost vertical. We typically guarantee down to about 10 parts left in the bin - but often we empty the bin completely anyway. In projects where the bins are rather small, we sometimes install a "shaking" mechanism to move parts away from unfortunate positions. In such cells we can always empty the bin.

* Cycle time is often an important factor. As mentioned by one of the other posts, we have the option of using sensors not mounted on the robot arm. In that case, the data is acquired while the robot is delivering a part and thus not obstructing the view into the bin. In other cases however, the price is more important than the cycle time. Here it is a better solution to mount the sensor on the robot arm because it is a more cost efficient solution. The reason we don't add two cameras is that the unit will become to large which means that we can't get into the corners of the bin. Finally, I can add that when we mount a 3D scanner on the robot arm (the alternative is a 2D camera), we only need to acquire one image so there is no extra cycle time added for moving the robot to a second position.

We will in the fall add some new videos to our homepage which will show some more impressive bin-picking. It has been planned a while but is taking some time because we need approval from the customers to post videos of the bin-picker cells for their products. So keep an eye on the home page during the fall.