ORACLE - Our First RobotORACLE - Our First RobotExploring Limits
With ORACLE shearing, exploration could begin. Well, almost. The first successful shearing test was an important milestone and the next was to repeat the results obtained by CSIRO. Only then would we be exploring unknown domains. It was one thing to shear a couple of blows, quite another to shear the back and side of a sheep reliably, at any speed.
High speed shearing tests would have to await the new cutter on the surface follower mechanism being assembled in the workshop, but there was plenty to get on with. There were countless small frustrations to overcome. The trial records abound with the phrase 'trial abandoned due to.....'.
A loose screw fell out and the capacitance sensors fell off. One of the 20 or so wires to the manual control box broke, and Rob Greenhalgh needed a day's work to locate the break and repair the cable. Wool became snagged in the wrist mechanism which had to be partially dismantled to remove it. A simple modification was needed to prevent a recurrence, but this meant another two days wait.
The cutter mechanism was particularly unreliable. Cutter problems would be with us for years to come, and had we known that, we might have fixed the problem there and then. But we wanted to avoid cutter problems. We had no wish to become involved in research on cutters as it seemed that other experts knew far more than we did. So we just dealt with the problems on an ad hoc basis. Disassemble, regrind, reset and reassemble. A day's work.
The rotary yaw actuator stiction (combination of friction and sticking, often causing jerky movement) problems which held us up for so long (p. 53) were just beneath the surface. From time to time they re-appeared, just to disappear frustratingly. Again, the new follower mechanism would provide a more permanent solution so there was little point in trying to fix the current one - we just persevered.
Another seemingly trivial problem - the potentiometer on the drive belt which measured the traverse position of the carriage seized. This puzzled us - we looked at a new replacement and the shaft was not a tight fit. There was a flexible coupling so that slight misalignments could be accommodated. So with a shrug, we just replaced the potentiometer and presumed that it wouldn't happen again. It did, several times.
All these problems were transient, fleeting, slight blemishes which we could deal with, or just put down to a hurried fix here, a patch there. Frustrating - yes but not a real concern. Soon we began to see the more fundamental difficulties which we had expected in one way or another, and it wasn't long before there were some unexpected ones too.
Our first shearing blows ran along the firm muscular side of the back, just a little way down from the backbone. The object was to discover how well we could control overlap without any explicit knowledge of the surface beforehand. Each blow was defined by:
- a starting position obtained by moving the robot under manual control until the cutter was just above the desired landing point on the sheep,
- a surface normal vector obtained by using the robot to measure another point further off the surface at the landing point,
- a direction vector parallel to the sheep, and,
- the desired shearing distance along this vector.
Once the first blow had been shorn, the next blow was planned automatically alongside the first blow with a given degree of overlap. The surface normal at the landing position was estimated from cutter attitude information remembered from the first blow. In effect, this was a strategy for a blind shearer with perfect motion recall. The results were surprising. A blow repeated by the robot would follow the same path on the sheep skin. With just a nominal amount of overlap programmed in the computer, wool would be neatly shorn with just wisps left between successive blows. More amazing though was the sight of the wisps remaining after the shearing pattern was run a second time over the shorn skin, even though the sheep had wriggled a bit. Clearly, blow positioning on the sheep was repeatable, at least under these conditions.
The sheep kept remarkably still, most of the time, but when they did struggle, there were some amusing and unexpected results. First, the sensitivity of the follower was surprising, even to us. As a sheep writhed and wriggled, more to relieve discomfort after lying still for a long time waiting for us to get moving than from any pain, the cutter rode up and down as if glued to the skin. Some sheep were genuinely ticklish. But then, when the cutter was not working properly, it caused some discomfort as it pulled the wool rather than cutting it. At other times, the cutter would rhythmically rise and fall, following the breathing of the sheep.
We replayed the robot movements on a screen - fourteen images per second. Sheep movement due to breathing showed as a series of small periodic ridges; struggling showed as jumbled chaos, a jagged disruption of an otherwise orderly record. Remembering that we had programmed each blow to follow the trace left behind by its predecessor, you can now visualize the consequences. The first blow would appear smooth enough; apart from a brief heaving and wriggling, the sheep would remain calm. But some distance along the next blow, with the sheep still, the robot heaved and wriggled, following the trace precisely inscribed in its memory! After we realized that our software was performing so faultlessly, it was amusing to watch.
Once we had demonstrated the basic mechanism for planning blows automatically, we attempted to implement our continuous shearing pattern. Instead of stopping at the end of one blow, and repositioning to start the next blow next to the first landing point, we turned the cutter on the sheep and followed the trace of the first blow backwards. This worked well for two or three blows, but as we extended the pattern down the side of the sheep we discovered some problems.
The consistency of the sheep flesh varies enormously from one part to another. We built up our confidence on the firm muscular rump quarters only to lose it on the soft wall of the abdomen. Here, when the sheep relaxed, the skin transformed into a loose membrane containing fluid, losing all its resilience. As the comb nosed in, seeking electrical contact, the skin retreated. As the capacitance sensors reacted to the warm closeness of the bulging membrane behind, the computer correctly pitched the cutter forward to dive the points deeper down. Then, when the sharp prongs finally made contact through the dense side wool, the comb would start to rise out again. But now the wool clung to the teeth, holding the skin in place, the comb rising higher and higher. Finally the skin fell away, and the comb stopped, and started to plunge in again. With hindsight, it is easy to understand what was happening; at the time, we could only be thankful that the sheep was completely unharmed. The only effect on the sheep was the rough and ragged swath of half shorn wool.
Part of the problem lay with the sheep support. The animal was comfortably supported at the brisket between the front legs and the crutch at the rear. The 'tummy' hung from the ribs and backbone. When the animal relaxed, this could sway some distance either side - it clearly had to be restrained. Roy added more support with webbing between two bars shaped to hold the bulging abdomen. Part of the problem lay with sensing.
We attempted our original unidirectional shearing pattern, with slightly better results. The continuous shearing pattern was retained for just the first two or three blows.
As our trials progressed, time passed and the wool on our flock of sheep grew longer and a new problem appeared. The cutter assembly was about 220 mm long, and with 120 mm high walls of wool on each side, its turning ability was severely impeded.
We decided to abandon our forwards and backwards shearing pattern. Each minor improvement served only to make the real difficulties all the more apparent. There was no room between the walls of unshorn fleece to turn the cutter, and the bluff shape of our cutter mechanism snagged the shorn fleece and stopped it from smoothly peeling away. After two or three blows, we had to stop shearing as the cutter head became lost in a growing bundle of tangled locks of fleece.
It was time to think again.
At least we had proved that memorizing the blow paths would allow successive blows to be shorn with minimal overlap. So we returned to David Henshaw's approach, shearing down the side of the sheep. We had the advantage, of course, that ORACLE could shear the opening blows along the top of the side, opening up a clear swath to start the downward blows.
And we were handsomely rewarded. Time after time, almost the entire side of the sheep was cleanly shorn, albeit slowly.
And so, by the end of the year, we had reached the stage of shearing nearly all of the back and one side of the sheep, using only a notional starting point, an initial shearing direction and a set of 'fencing' planes and of course the sensor inputs to guide the comb. In December and January 1980 the new follower mechanism was installed on the robot and we could begin to shear faster. Soon we were to discover the advantages of encoding more sheep knowledge, refine our motion control techniques, and improve our sensing. Over the following years, we often looked back at the photographs and videotapes recording those five months of shearing and wondered whether all the work in between had achieved anything. It had of course, but sometimes it seemed that we couldn't shear nearly as well.
Epilogue
ORACLE served us well for six years of shearing trials. Its reputation
as the new breed of Australian shearer spread, as figure 2.14 shows. Apart
from its one lunge at the Vice-Chancellor it was remarkably well behaved
in all that time. The oil hardly ever leaked. It ran off its track once
- a runaway software error. The results of our five hundred shearing trials
were coalesced into its successor in 1984 and 1985 (see chapter 8). Our
decision to mount the hydraulic connections on the outside was a mistake
- we never made significant changes. But our decision to place the wiring
on the outside was a good one - we had many wiring failures and had much
to learn about designing wiring for flexure and vibration.
|Tour Page|James Trevelyan's
Page| UWA Home page|
Web site design: James Trevelyan jamest@mech.uwa.edu.au
September 1997