G64 P and Q - Strange behaviour

Hey guys, I've encountered some odd behaviour with the naive cam detection/path blending and I'm curious as to whether this is expected behaviour and my understanding is wrong, or this indicates something else.

Context:

• Path is long, smooth curves (consisting of many small straight segments) with a horizontal lead out and a 2mm stepdown (see 'baseline' image below for an overview)
• Images are taken from the backplot, viewed along the Y axis (in images Z+ is up and X+ is right)
• Baseline setting of G64 P0.5 Q0.5 (machine units mm)
• System is an ethercat servo system, with accelerations set in linuxcnc far below what the drives and mechanics are able to do (ie the drives can comfortably keep up with LCNC's instructions)
• X axis is much faster top speed but slower accel than the other two axis.
• In the images the long gentle curves are ~1.8m and are run at 10m/min (machine cuts foam) and the total Z travel for the curve is only 50mm. The Z axis can comfortably keep up with X in these cuts
• There is no Y component to the motion in these shots.
• Rerunning any version results in identical results (ie hit start again and the traveled path on backplot lines up perfectly no matter the zoom level) suggesting it's not following error or similar, ie the trajectory planner is more likely the cause than drives/machine issues

Observed behaviour:

1. Not touching path:
   • With G64 P0.5 Q0.5 set the traveled path on the backplot never touches the gcode path... for the entire 1800mm curve length. See images 'Baseline' and 'Baseline 2'.
   • The curve starts low, Z=20, peaks at z=70, then drops back to approximately the same height. Even at the very long, near-horizonal section (ie where Z motion is near zero), the traveled path does not touch the gcode path, instead oscillating nearer and further away.
2. Setting Q to low values seems to impact path blending:
   • See the remaining images (q0.1, q0.01, q0.001)
   • The P value for all images was 0.5
   • Very low Q values seem to change how path blending is done, ie LCNC stops cutting corners and acts as though the P value has been reduced to the same value as Q. Image Q0.001 shows this the clearest, where it looks like exact path mode rather than blending, despite the path blending variable still being 0.5
   • At Q=0.01 there's the tiniest corner cut (ie significantly smaller than P0.5 would theoretically produce) and at Q0.1 it looks more normal but the path on the long curve struggles to touch the intended line

So observation 1: shouldn't the traveled path touch the gcode path periodically?
and observation 2: shouldn't setting a low value for naive cam detection not impact the path blending's behaviour for smoothing corners?

Final, slightly out there thought: could velocity feedforward or other similar cia402/ethercat elements impact this? System is currently running ethercat drives in cia402 position mode, without any PID on lincnc's end, could that be the cause?

Thanks in advance for people's insight

from my current research, the Q value determines when to collapse connecting line segments based on how far their joint is from being collinear with their outer endpoints. this reduces the number of small segments in the trajectory planner. the P value determines how far away from the joint of 2 line segments the path can deviate. this is the rounding of corners you would experience. so if Q value shrinks and more line segments are present, this will force the path to more closely follow the existing Gcode lines by way of creating more points for which the P value is applied to. we also know that the trajectory planner must touch each line segment at least once, so even if the P value is set very high, the TP cannot deviate by very much give such a short distance between two consecutive contacts on each small segment.

as the Q value grows and more of small segments are merged into longer segments, the distance and amount of freedom between the two segment touch points and their joint increases allowing for the ability to achieve larger P value settings.

here is a rudimentary graphical representation as I understand it.  if I have this wrong please someone correct me!

 

The path you see in the preview has no dependance on anything external to the software and settings.
The system tries to do the best it can within your G64 limits _and_ the acceleration and velocity limits in the INI.

You mention that the iNI acceleration and velocity are conservative relative to what the machine can really do. I think that you will see far better path following if you increase the INI values to be less conservative (especially the acceleration)

Hey guys, thanks for the swift and detailed responses; Lcvette your explanation and drawing helped clear some things up for me too.

The issue I'm having is what I'm seeing doesn't match those described behaviours (or I've misunderstood or missed something, in which case thanks in advance for your patience) in two key ways:

1. "..the trajectory planner must touch each line segment at least once" - I'm seeing over a meter of travel, with around a thousand segments, where the path doesn't touch at all (see 'not touching points' image for hasty mspaint diagram) unless Q is set extremely low (0.5 = no touching/intersection, 0.01 does touch).
   • Could this be due to some sort of averaging of the path? ie when niave cam detection acts across all segments in a long convex curve it averages the curve out towards flat?
2. Low Q + lots of points = less freedom for P to act - This is where I'm more sure that the behaviour is odd: at the end of each curve is a 30mm horizontal lead out, 2mm step down, then 30mm lead in. When Q is set very low (ie to more accurately trace the curved paths) the lead ins and lead outs also get traced more accurately, as if the P value is ignored (see 'negated P' image).
   • These motions are four points: (I just checked the gcode to be sure)
     • end of curve/start of lead out,
     • end of lead out +30mm in X
     • stepdown -2mm in Z
     • lead in/start of next curve -30mm in X

For a bit of context: This is a specialty machine that needs to work with a specific cam (that is still under development so I've been able to get a feature put in to reduce point density, though my results are similar with said feature), I need to be able to use path blending to smooth out the early stage transition moves (linear rectangles for now, asking for tangential extension curves but it's not coming anytime soon) while still reasonably accurately following the curved paths, hence the need for a low Q and a relatively high P. If I could get it to at least touch each segment post-Q point-decimation it would work perfectly, I'm just unable to get it to touch most line segments for reasonable Q and P values.

andypugh: Y and Z have very little work to do to keep up with X (long shallow curves) but they've got plenty of accel left on the table so I'll crank their accel right up and see what I get.

Cheers for your time guys.

Additional question/thought: is there a way to export/log the traveled path (ie the red line in backplot)?
 

I believe that if you set Q=0 it will turn off the Naive CAM Detection and the path should then touch a point on every line segment, if that is what you are wanting.

PS  If Q is very low or zero, then yes, there will not be room for P to use it's full blending tolerance if line segments are short enough to not allow the full tolerance and still touch the segment.  Pretty much if any segment is less than twice P it will limit the amount of blending regardless of the length of adjacent segments.  Having a large P and very small Q will likely result in mixed tolerance results where parts of the code with short segments will have a close tolerance but any corners with adjacent long segments will be significantly rounded.

That's the weird thing, large P and low Q should round corners on segments where points are sparse but it's not, see the 'negated P' image in my previous post for a diagram of what I'm seeing. Basically with P0.5 and Q0.5 the lead ins and lead outs (large rectangular moves) are being rounded as expected, but with P0.5 and q0.01 they're not, instead it's sticking within Q tolerance of the points.

Unless I've missed something (which is entirely possible of course) everyone (including me) agrees on what P and Q should do in this situation but my observations don't match that, any ideas on what to check next?