Skip to content
noggin

Perception and appeal for the next Mech's face

Researched 2026-11-02. Purpose: test #51’s face choices against perception research. The choices are linked eyes, a big head on a small base, Pure White eyes with black pupils, separate lids, and gestures timed for a few servos. Builds on 2026-10-31-natural-eye-motion.md, 2026-11-01-blink-kinematics.md and 2026-11-01-toon-eyes.md. Their timing numbers (Williams, Thomas & Johnston, Lasseter, blink kinematics) are not repeated here.

Tags: [A] = abstract only (paywalled or not reached), [P] = preprint or conference abstract, [I] = my inference or arithmetic. Geometry used: eye centres 59 mm apart (model/geometry.py EYE_R), viewer 0.5–1 m away at a desk (inferred).

Sources:

  • Baby schema raises cuteness and the urge to care [G]. High schema means a round face, high forehead, large eyes and a small nose and mouth.
    • Manipulations stayed within ±2 SD of 40 real infant faces.
    • Eye width / face width averaged 18.5% (SD 0.9), so the high-schema end is at most ~20.3%.
    • Cuteness: F(2,59) = 137.3, p < .001, with high > unmanipulated > low; no sex difference in cuteness ratings.
    • [G] did not test values past the natural range, so it says nothing about cartoon proportions.
  • Head: perceived cuteness falls as head shape ages (cardioidal strain) [Al]. Teddy bears changed over 1903–1985 toward a larger forehead and shorter snout [HB], from only 15 bears.
  • Mech, inferred: two caps “up to touching” on 59 mm centres give eyes ~59 mm wide each. On a ~150 mm head that is ~40% of face width per eye, twice the infant extreme [I]. That is a supernormal cartoon value: it works on a stylised face and turns eerie on a realistic one (§4).
  • White sclera exists to signal gaze [KK]. Humans have the largest ratio of exposed sclera among primates, and it is unpigmented. Gaze reading treats the darker region as the part doing the looking [R]: dark on light reads accurately, reversed polarity reads poorly.
    • Morty and Bart’s black pupils on Pure White fit this.
    • Their small pupils leave a lot of white showing. A large visible white is a fear and alarm signal [Wh, amygdala response to fearful (larger) eye whites].
  • Lid position is the strongest single eye cue [LA]. Eye widening against narrowing explained 61.7% of the variance in perceived mental states, and the eyes still carried it against conflicting lower-face cues.
    • Widening reads as sensitivity: awe, interest, fear. Narrowing reads as discrimination: suspicion, disgust.
    • FACS surprise is sclera showing above the iris (natural-eye-motion §4).
  • Pupil shape (Morty jagged, Bart dots): I found no perception study. It is style.
  • The cone of direct gaze is wide and stable [GH]. People feel looked at across ~8–9° at 1 m (virtual head 9.3°, real head 8.1°), which is about ±4°. A real head at 5 m gave 3.9°.
    • The cone is far wider than acuity, whose difference thresholds are ~1°.
    • A 10° head turn pulled the real head’s cone only 0.74° toward the head direction.
    • [Ot]: the head acts on perceived gaze in two ways. It repels gaze through the eye region and attracts it as a coarse direction cue; together they combine linearly.
  • Vergence tolerance [LH22]: observers accepted ~3.5–4.5° of convergence and 2.5–3.5° of divergence as still “looking at me”. The exact point of being looked at was 0°, which the abstract calls “parallel eyes”.
    • I could not check whether 0° means truly parallel or converged on the observer.
  • Two-eye readings [LH24]: the more rotated, inward-turned eye sets the judgement. Its direct-gaze area is ~4°, against ~6° for the outward-turned eye.
  • Bias toward convergence [N]: observers see others’ gaze as more convergent and nearer than it is, especially downward gaze or under uncertainty.
  • Misalignment detection [La]: lay observers reliably detect a horizontal misalignment of 12.5 prism dioptres (PD), about 7.2°.
  • What linked eyes cost [I]:
    • Converging on a viewer at 0.6 m needs 5.6° (9.8 PD), at 0.45 m 7.5° (13.1 PD), at 1 m 3.4°.
    • Parallel linked eyes are therefore inside the vergence tolerance from ~0.8 m out. They sit below the misalignment threshold down to ~0.5 m.
    • The real loss is near targets: the Mech cannot convincingly fixate anything closer than ~0.5 m, such as its own base or a phone beside it.
    • It also loses the cross-eyed gag.
  • Fixed toe-in [I]: set the linkage to converge at ~1 m, 1.7° per eye toward the nose. Then 0.5–2 m stays within ±3.5° of the needed vergence, and it is also within [LH22]’s tolerance if their 0° means parallel.
    • At a 25 mm cap radius this is a 0.74 mm inward pupil offset. It can live in the pupil seats of the caps, with no gimbal change.
  • Pupil placement error [I]: 1 mm of pupil slip on a 25 mm cap is 2.3° of gaze, half the cone. A magnet-held pupil that can slide or seat loosely moves where the eyes look.

3. Animation principles on position servos

Section titled “3. Animation principles on position servos”
  • Anticipation and reaction, tested on a robot [TDJ] (PR2 videos, N = 273, 1–7 scales):
    • A visible “forethought” before acting raised appeal from 4.27 to 4.83, approachability from 4.54 to 5.05, and viewers’ sureness of their reading from 5.20 to 5.69.
    • It did not raise the accuracy of that reading (keyword match, p = .37).
    • A reaction to the outcome made the robot seem smarter: 3.86 against 4.72.
    • So anticipation buys likability and confidence, not legibility. The legibility comes from the gesture’s main pose.
  • Mapping each principle to a position servo [I, on TJ/L]:
    • Anticipation: a small commanded counter-move before the main move.
    • Overshoot and settle: command past the target, then back. Do not rely on the servo’s own overshoot, which is load-dependent.
    • Slow in/out: interpolated targets at 20 ms; lost on moves under ~3 ticks.
    • Arcs: start the two axes of a pair (eyes L/R + U/D, or pan + tilt) together but with offset easing, so the path curves instead of running straight.
    • Secondary action: lids follow vertical eye moves (Becker & Fuchs, natural-eye-motion §3), and the mouth follows head accents.
    • Overlap: the eyes start before the head on a turn. This is common animation practice; I found no primary source in scope.
    • Timing and holds: see toon-eyes §3–4 (the 1.67 s still limit, the 250 ms accent).
  • Motion can deepen the valley [M]: Mori’s smile slowed by half “turns creepy” and motion amplifies the peaks and valleys. [K] found that movement’s modulating effect is not consistently supported.
    • [I]: scale Intensity by amplitude and keep the timing. Slowed-down gestures are the risky knob.
  • The valley comes from mismatch, not from being a robot [K, SN, MD]:
    • Mismatch between realistic and unrealistic features is the best-supported road into the valley. The naive “more human = eerie near the top” curve is not consistently supported [K].
    • Abnormal eyes are eerie only on faces that are otherwise realistic [SN].
    • Atypical proportions disturb more on photorealistic faces, and a mismatch of eye size or texture with the face is especially eerie [MD].
    • [MR] found the valley in likability across 80 real robot faces.
  • Mori’s advice: aim for the first peak, moderate human likeness such as toy robots and bunraku puppets [M]. A flat-colour cartoon head on a visibly mechanical neck sits there [I].
    • Its risk is internal mismatch: a realistic eye (glossy, textured iris, lashes) in a matte toon face [I from MD].
  • Blink rate shapes impressions [Ta]: avatars were shown at 9, 12, 18, 24 and 36 blinks/min.
    • 18/min was friendliest, 9/min read as intelligent, 36/min as inactive.
    • The effect was larger on human-style avatars than on cartoon humans or animals.
    • The current thinking rate, one blink per 5–8 s (7.5–12/min), sits at the “intelligent” end.
  • Long blinks read as “got it” [Hö]: a 607 ms listener blink, against 208 ms, made speakers give shorter answers, though none noticed the difference. It works as an acknowledgement.
  1. Keep both characters fully stylised and consistent: flat matte PLA, no realistic iris texture, gloss or lashes on a toon face [MD, SN, K, M].
  2. Black pupils on Pure White eyes, never reversed or low-contrast [R, KK].
  3. Make the pupils keyed, not just magnet-held, to within ~0.5 mm, which is ~1° of gaze. A 1 mm slip costs half the 8° gaze cone [GH, I].
  4. Converge the linked eyes at ~1 m: 1.7° per eye, or a ~0.7 mm inward pupil offset on 25 mm caps [LH22, N, La, I].
  5. Hold the linked eye gimbal’s play to ≤ ~2° (half the gaze cone) where it sets eye contact [GH, I].
  6. Don’t stage gestures that fixate within ~0.5 m (the base, a phone). Look at the user or far away [I from LH22].
  7. Rest lid at or just over the top of the pupil, so no white shows above it. Show white above the pupil only for eureka and attention (widening = interest or awe), and narrow for thinking and wince [LA, Wh].
  8. Big head, high forehead, small nose and mouth. The cranium electronics should push the forehead up, not the eyes down [G, Al, HB].
  9. Anticipate before main moves and react after outcomes; expect appeal, not legibility, from it. Legibility needs a clear main pose [TDJ].
  10. Scale Intensity by amplitude, not by slowing the gesture [M, I].
  11. Blink near 18/min when engaged and ~9/min when thinking. Avoid ≥ 36/min except in flutter gags [Ta].
  12. End attention and eureka with one long blink (~600 ms) as the “got it” beat [Hö].
  • No #51 decision is beaten by this evidence. Linked eyes, the big head on a small base, white eyes with black pupils, separate lids and caps up to touching are all supported or neutral.
  • Linked eyes, cost bounded: gaze reads as direct from ~0.5–0.8 m out (§2), and linkage only loses near targets and cross-eye gags.
    • Proposed addition (no servo or limit change): a fixed ~1 m toe-in, either in the gimbal’s link geometry or in the caps’ pupil seats (practice 4).
  • Gap: #51 ranks gimbal candidates on simplicity, torque, fit, leads, tip-over and pack access, with no perceptual criterion.
    • Proposed: add “gaze error ≤ 2°, including play and pupil seating” as a pass/fail before the ranking (practices 3, 5).
  • Outside #51: the glossary’s Rest (“servos holding so the mech is still”) conflicts with Thomas & Johnston’s 1.67 s still limit (toon-eyes §1) and with [Ta]’s blink effect.
    • Proposed: a blink every 3–6 s at Rest, if servo noise and power allow. That is Cansin’s call.

Unverified: [A] items rest on abstracts only.

  • [LH22]’s “0° = parallel” geometry is unchecked.
  • The “50% enlarged eyes” in [MD] came via a press summary, so I left it out.
  • No source tests cuteness past natural proportions, or perception of pupil shape.
byYellow PineYellow Pine

noggin v0.2.26-6-g09bf9b7 · 09bf9b7