Historical page (2026-11-08). The Monte Carlo was run before V6 (2026-11-14) and V7 (2026-11-17), so it does not describe the current robot. See the current robot, V7.

Monte Carlo of 2,000 balls per ball type: calibrated shot scatter

Updated 2026-11-08. Model results with assumed parameters, not measurements. The earlier 2,000-ball run used the flow model's own shot scatter, since shown to be 5-20x too large; those numbers are superseded and shown here as "old".

Headline. With the calibrated scatter (elevation sd 1 deg, azimuth sd 1 deg, speed sd 2%) the design-of-record FLOWER funnel (65 deg, mouth +55 mm) delivers 91.0% of loaded POLLEN and 77.9% of loaded NECTAR (98.8% and 96.1% of the balls that were actually launched). Calibrating the scatter changed the FLOWER numbers by only 0 to +10 points; the largest FLOWER loss is the intake (150 POLLEN and 374 NECTAR of 2,000 never get in). The HIVE is what the calibration changed: 63.3% POLLEN and 66.5% NECTAR with the speed tuned per distance, against 12.8% and 9.2% before.

FLOWER: old scatter against calibrated scatter (per 2,000 loaded balls, 95% CI)

BallFunnelOld scatter (superseded)Calibrated scatterCalibrated, given launchedCalibrated failure modes (tube and drop, of 2,000)
POLLENCurrent funnel (53 deg, as built)77.8% (75.9-79.5)87.8% (86.3-89.2)95.3%tube: fell back out of mouth: 80
drop: bounced out: 4
tube: blocked at mouth rim: 2
tube: stuck inside: 1
POLLENRe-aimed 70 deg / mouth +45 mm90.9% (89.6-92.1)91.8% (90.5-92.9)99.6%tube: fell back out of mouth: 4
drop: bounced out: 3
tube: stuck inside: 1
POLLENRe-aimed 70 deg / mouth +55 mm90.6% (89.3-91.8)91.9% (90.6-93.0)99.7%tube: fell back out of mouth: 4
drop: bounced out: 1
POLLENDesign of record 65 deg / mouth +55 mm91.0% (89.7-92.2)*91.0% (89.7-92.2)98.8%drop: bounced out: 21
tube: stuck inside: 1
NECTARCurrent funnel (53 deg, as built)0.0% (0.0-0.2)0.0% (0.0-0.2)0.0%tube: blocked at mouth rim: 1621
NECTARRe-aimed 70 deg / mouth +45 mm74.2% (72.2-76.1)75.7% (73.8-77.5)93.4%tube: fell back out of mouth: 96
drop: bounced out: 10
tube: stuck inside: 1
NECTARRe-aimed 70 deg / mouth +55 mm77.5% (75.6-79.3)75.2% (73.3-77.0)92.8%tube: fell back out of mouth: 100
drop: bounced out: 16
tube: stuck inside: 1
NECTARDesign of record 65 deg / mouth +55 mm78.9% (77.1-80.6)*77.9% (76.0-79.7)96.1%drop: bounced out: 38
tube: fell back out of mouth: 25

* The design of record was not in the first study; re-run here with the old (flow-model) scatter on the same chain runs. Bore 95.6 mm (the 103.2 mm sleeve section is below). Intake and feed losses are the same in every row: POLLEN 150 + 7, NECTAR 374 + 5 per 2,000.

Design of record 65 deg / 55 mm: sensitivities (calibrated, per 2,000 loaded)

CasePOLLENNECTAR
Azimuth centred on 0 (instead of the flow-model offset)91.5% (90.1-92.6)77.5% (75.6-79.2)
Bore 98 mm90.3% (88.9-91.5)77.6% (75.7-79.4)
Bore 103.2 mm (sleeve section)76.3% (74.4-78.2)68.7% (66.6-70.7)

103.2 mm over the whole tube costs about 9 points for NECTAR and 15 for POLLEN (balls bounce out at the drop). The design only has 103.2 mm in the 129 mm vertical sleeve section; that is modelled in the next section and costs about 0.4 points.

Design of record with the real v6 scarf, the 103.2 mm sleeve and the built flats (calibrated, per 2,000 loaded, single blind pass)

Label. Every rate below is a single blind pass: each ball is approached once, with no retry and no wall. It is a worst case for the pickup, not a prediction of a match. The realistic table further down says what changes when the driver keeps driving, uses walls and retries.
Funnel / ball pathPOLLENNECTAR
per loaded (95% CI)given launchedper loaded (95% CI)given launched
Generic study tube (scarf 16 deg), 95.6 mm bore: the earlier 65/55 run91.0% (89.7-92.2)98.8%77.9% (76.0-79.7)96.1%
Real v6 scarf (A_ 61.6,227.8; B_ 185.9,212.8: 18.1 deg), 95.6 mm bore92.0% (90.8-93.2)99.9%77.0% (75.1-78.8)95.0%
v6 scarf + the real 103.2 mm sleeve section in the ball path (z 244-373 only)92.1% (90.8-93.2)99.9%76.6% (74.7-78.5)94.6%
  same, tube axis at the CAD x = -7 mm, round funnel92.0% (90.8-93.2)99.9%75.4% (73.5-77.3)93.1%
  same, with the built flats (98 mm wide front)92.0% (90.8-93.2)99.9%75.4% (73.5-77.3)93.1%
103.2 mm over the whole path (old worst case, not the design)83.2% (81.5-84.8)90.3%67.2% (65.1-69.2)82.9%

Is any of the ball path 103.2 mm? Yes, but only the sleeve section. The sleeve bore is 103.2 mm and the riser (OD 102.8, ball bore 95.6) only enters its top 19 mm when the tube is up, so the ball travels through an empty 103.2 mm bore from z 244 to z 373 (129 mm, vertical), then steps down to 95.6 mm in the riser, arch and spout. Everywhere else (funnel, bend, riser, arch) the bore is 95.6-96 mm. Modelled as 103.2 mm over that 129 mm only, the cost is about 0.4 points for NECTAR and none for POLLEN; the earlier 9-15 point loss was for 103.2 mm over the whole path, which the design does not have. Not modelled: the 3.8 mm ledge where the ball meets the riser bottom (give it a chamfer).

Flats. The funnel front is trimmed to a flat-sided band 98 mm wide (|x+7| > 49 removed over the 40 mm funnel). In the model no ball ever touches that part of the wall: balls leave the shooter within 17 mm of the tube axis, and even flats only 20 mm wide give identical results. Flats against no flats: no change (POLLEN 92.0% against 92.0%, NECTAR 75.4% against 75.4%, same chain states). The 2 mm axis shift from x = -5 (chain study) to x = -7 (CAD) moved NECTAR down 1.2 points (more fall-backs out of the mouth), within the noise.

Intake options (bench alternatives; first 300 chain runs per ball, paired with the plan; per loaded ball, single blind pass)

OptionPOLLENNECTAR
picked uplaunchedend-to-end FLOWER (95% CI)vs plan (points)picked uplaunchedend-to-end FLOWER (95% CI)vs plan (points)
Plan: bar -16, float 60 (grip as assumed)92.7%92.3%90.8% (86.9-94.5)-83.7%83.7%78.0% (73.2-82.1)-
Plan, roller grip -20%89.0%88.7%88.4% (84.4-92.1)-2.3 (-5.4..+0.6)56.0%56.0%52.1% (46.4-57.7)-25.9 (-30.8..-20.6)
Plan, roller grip +20%95.0%94.7%94.2% (91.1-96.9)+3.4 (+0.9..+6.3)93.0%93.0%87.8% (84.4-90.8)+9.8 (+6.3..+13.7)
Spare: bar -20, float 6089.7%64.3%64.1% (58.3-69.9)-26.6 (-33.1..-20.7)96.7%96.7%90.8% (88.1-93.2)+12.7 (+8.8..+17.0)
Spare, roller grip -20%89.0%79.7%79.2% (74.4-83.9)-11.5 (-16.3..-7.2)87.0%87.0%82.7% (78.8-86.3)+4.6 (+1.2..+8.1)
Spare, roller grip +20%88.3%46.3%47.3% (41.6-53.3)-43.4 (-49.7..-37.4)99.3%99.3%92.8% (90.9-94.2)+14.7 (+10.5..+19.4)

Roller grip means the assumed rubber-on-ball friction of the fixed bar (scaled 0.8 or 1.2 on every run). Lost NECTAR are almost all balls pushed ahead of the intake, stuck against the bar and Gecko row: they are lost when the bar grips less than the Gecko (roller friction / Gecko friction below about 1.1). The spare -20 bar helps NECTAR (+12.7) but strands POLLEN at the deck entry (about 1 in 4 enter and jam), so POLLEN falls 27 points. Bench the plan bar -16 first.

Pickup scenarios beyond one blind pass

BarScenarion POLLENPOLLEN picked upPOLLEN end-to-endn NECTARNECTAR picked upNECTAR end-to-end
Plan bar -16Single blind pass (the study's default)20090.5% (85.6-93.8)89.8% (85.3-94.1)20082.5% (76.6-87.1)78.1% (72.6-83.2)
a: robot keeps driving 0.5 s after capture20090.5% (85.6-93.8)89.8% (85.5-94.3)20082.5% (76.6-87.1)77.6% (72.1-82.9)
b: ball 0-30 mm from a wall, robot drives into it (60 N push cap)20012.0% (8.2-17.2)12.8% (8.0-17.8)20046.5% (39.7-53.4)44.2% (37.3-51.4)
c: ball in a corner (two walls)2000.0% (0.0-1.9)0.0% (0.0-0.0)20013.5% (9.4-18.9)0.4% (0.0-1.2)
d: one re-aimed retry after a lost ball20099.5% (97.2-99.9)99.0% (97.7-99.9)20087.0% (81.6-91.0)82.0% (77.1-86.3)
Spare bar -20Single blind pass (the study's default)5086.0% (73.8-93.0)62.8% (49.7-76.5)10096.0% (90.2-98.4)88.8% (84.3-92.4)
a: robot keeps driving 0.5 s after capture5086.0% (73.8-93.0)69.4% (57.3-81.7)10096.0% (90.2-98.4)89.6% (84.8-93.1)
b: ball 0-30 mm from a wall, robot drives into it (60 N push cap)5022.0% (12.8-35.2)18.2% (8.3-30.7)10036.0% (27.3-45.8)31.8% (22.7-40.6)
c: ball in a corner (two walls)500.0% (0.0-7.1)0.0% (0.0-0.0)10012.0% (7.0-19.8)1.8% (0.0-5.5)
d: one re-aimed retry after a lost ball50100.0% (92.9-100.0)73.6% (61.5-85.5)10097.0% (91.5-99.0)89.5% (85.0-92.9)
Plan bar -16b': same wall, gentle push (20 N cap)10076.0% (66.8-83.3)75.3% (66.1-82.8)10083.0% (74.5-89.1)77.5% (70.7-84.0)

Open floor with one re-aimed retry is the only scenario that clearly beats the blind pass (the retry re-aims x and angle, but the ball and grip are the same, so friction-limited NECTAR losses mostly repeat). Driving on after capture changes nothing. The wall result depends strongly on how hard the robot pushes (a gentle push helps, a hard push crushes the soft ball); the corner is 0% because a corner ball sits outside the 104 mm lane and bar (it needs a different technique). These are untested assumptions; n is 200 per cell for bar -16 (100 NECTAR and 50 POLLEN for the spare).

Realistic mix (assumption: 70% open floor with one retry, 25% against a wall, 5% in a corner), per loaded ball

NumberPOLLENNECTAR
FLOWER, single blind pass (worst case, plan bar)92.1%76.6%
FLOWER, realistic mix, gentle wall push88.1%76.8%
FLOWER, realistic mix, hard wall push72.5%68.5%
HIVE, single blind pass58%54%
HIVE, realistic mix, gentle / hard wall55.9% / 45.8%54.3% / 48.2%

HIVE per loaded ball = launched share x HIVE given launched (63.3% POLLEN, 66.5% NECTAR, calibration study). The mix weights are my assumption; there is no match data. The model is not validated against the real robot: goBILDA's stock robot is seen picking up balls in its video, but the notes hold no pickup counts to compare, so the real pickup rate is unknown.

HIVE, from the calibration study (not re-run)

BallOld (flat window, old scatter)New (3-D CELL, calibrated scatter)Per loaded ballStock 6.0 m/s, no tuning (old / new)
POLLEN12.8% (11.4-14.3)63.3%58%2.1% / 3.9%
NECTAR9.2% (8.1-10.6)66.5%54%4.0% / 11.5%

1.5-3.0 m, bearing 0-45 deg, speed tuned per distance and bearing (6.0-9.7 m/s); the stock speed only reaches the CELL at about 0.7-1.7 m. New figures include the corrected 3-D CELL geometry as well as the calibrated scatter.

What is not proven

Source: sim_study/mc2000/RESULTS_calibrated.md, mc_cal.py.