The policy now plays the G-major scale end to end. Not well yet. We know why, and we need more GPU time to fix it.
G-major scale · plays all 47 s · 16 / 70 notes frettedHouse still 8 / 8 · 51 % of notes sound right5 GPU-hours per experiment · 2 a day
Mean policy · 10 worlds · guitar offset ±5 mm / ±2° · pushed twice with 40 N · whole piece from t = 0 · House video carries the sound the policy played.
Since last week
5.4 → 8 / 8
House chord windows fretted
new
second piece: G-major scale, 47 s
new
score by sound, not only by force
1 · It plays
G-MAJOR SCALE · frets 2–7, all six strings · 47 s · 70 notes
Left: the reference. Right: the policy. Green = the note to press, red = where the finger pad landed.
16 / 70
notes fretted (was 10)
5 %
notes sound right (pitch + timing)
9 mm
guitar motion
HOUSE · 21 s · same recipe, unchanged
Perturbed world. Sound on.
8 / 8
presses
51 %
notes sound right (pitch + timing)
17 mm
guitar motion
2 · How well: four metrics
Metric
G-major scale
House
Target
Pressed what it should
16 / 70
8 / 8
all
↳ Sounds right: share of notes with the right pitch, starting within 50 ms of the tab counted both ways: of the notes played, how many belong; of the notes written, how many were played (the F1 score of the two)
5 %
51 %
≥ 80 %
Touched nothing it should not hand moves where no non-staying finger pushed a string above 2 N (a player's light brush)
12 %
43 %
100 %
Guitar stays put (max motion)
9 mm
17 mm
≤ 8 mm
Force in the human range · mean / peak (p99) while pressing
5.1 / 8.5 N
5.8 / 8.2 N
peak 6–19 N
How "sounds right" is scored
Physics forces
ours
→
Forces → played notes
our rules: pluck, fret, mute
→
Notes → audio → compared with the tab
Ruocheng's package, unchanged
3 · What is not solved yet
Left hand: the whole hand sits one string too far to the bass side
Right hand: rushes and misses (both pieces)
33 %
notes 50–120 ms early / late
16 %
notes not plucked
29 %
House plucks missed
Reference: right place, but the robot's pads are wider than a player's
28 / 70
fretted when played open-loop
5 mm
ring / pinky slide when pressed
4 · What it takes: iterations
5 h
GPU per experiment
2 / day
experiments on the one 4090
~10
variables tried this week
Numbers: pieceaccept / rescore_groups on eval_perturb10o (s197 ck12, s231 ck12); note score from 17_music_eval. Videos: kinematic replays of stored evaluation states with audio synthesised from the played events. Living record: notes/GUITAR_DEX_PROJECT.md, notes/daily/D05–D09.