Gantry

Evaluation report, Menlo Robot Hand Poses

We fine-tuned the same vision to joint model twice, once on Menlo robot-hand-poses-train as released and once on an edited copy with the 2,293 poses no single tendon can hold removed, and the edited copy scored 5.5 points higher across 840 held out predictions.

Fine-tuned on full
26.3%
Held out joint group hit rate with every pose kept.
CI 23.4 to 29.4 · n=840
Fine-tuned on edited
31.8%
Same run with the 2,293 poses a Shadow Hand cannot be commanded into removed.
CI 28.7 to 35.0 · n=840
Lift
+5.5 pp
From training on less data.
95% range +1.1 to +9.8
Removed
20.8%
2,293 of the 11,000 poses read, each asking one tendon to hold its two joints more than 1.20 rad apart.
worst pose asks for 1.5638 rad · 89.6 deg
Frozen joints
2 of 24
lh_WRJ2 spans 0.0006 rad and lh_WRJ1 spans 0.0165 rad across every pose read.
the other 22 span 0.53 to 2.08 rad

Success rate by task

fine-tuned on full Menlo Robot Hand Poses fine-tuned on edited Menlo Robot Hand Poses

The seven groups are the joint families the model has to emit, the four tendon coupled finger pairs, the thumb chain lh_THJ5 to lh_THJ1, the two wrist joints and the four abduction joints, and a held out pose counts as a hit for a group when every joint in that group lands within 0.1 rad. A same size random cut scored 26.6%, so the gain is from which poses were removed, not from removing poses.

Gain per task

The groups that gain most are the coupled finger pairs, which is exactly where the cut lands. The wrist goes the other way and loses three, which is what a frozen output should do, since no cut can teach a model a constant it already emits. At 120 poses a group most single groups do not clear zero on their own and the chart says so.

What the edited version had removed

Measured on the release. The Shadow Hand drives 24 joints with 20 actuators, because four of them are fixed tendons, lh_FFJ0 lh_MFJ0 lh_RFJ0 lh_LFJ0, each pulling one finger's J2 and J1 together, so the split between those two joints is not a thing a controller can set. In the 11,000 poses we read the four pairs are labelled as if they were free, with correlations of -0.083, -0.083, -0.084 and -0.067 where one shared tendon would give close to +1. 2,293 poses, 20.8% of what we read, ask a single tendon to hold its two joints more than 1.20 rad apart, and the worst asks for 1.5638 rad, a middle knuckle at 1.5593 with the fingertip at -0.0045. Those are targets no hand can be scored against, and the rate in the released test split is 22.0%, so the eval carries them too.

Removedposesworst split
one coupled pair commanded over 1.20 rad apart2,1041.5638 rad
two coupled pairs commanded over 1.20 rad apart1841.5272 rad
three coupled pairs commanded over 1.20 rad apart51.4970 rad
Removed2,2931.5638 rad
Kept8,7071.1999 rad

The removed media

Measured on the release. Each pose in this release is one 1280x1280 render and there is no motion anywhere in the repository, so these are stills. Every caption carries that pose's own two coupled joint values and the split between them, in radians and in degrees.

train row 67399
train row 67399removed
pose_043642, ring finger, middle knuckle fully bent with the fingertip dead straight
lh_RFJ2 1.5593 · lh_RFJ1 -0.0045 · split 1.5638 rad · 89.6 deg · widest in the 11,000 read
train row 68354
train row 68354removed
pose_043081, middle finger, the same shape one tendon cannot pull
lh_MFJ2 1.4976 · lh_MFJ1 -0.0296 · split 1.5272 rad · 87.5 deg
train row 63404
train row 63404removed
pose_028173, little finger, fingertip bent with the middle knuckle straight
lh_LFJ2 -0.0306 · lh_LFJ1 1.4967 · split 1.5273 rad · 87.5 deg
train row 56504
train row 56504removed
pose_026628, first finger, fingertip bent with the middle knuckle straight
lh_FFJ2 -0.0275 · lh_FFJ1 1.4933 · split 1.5208 rad · 87.1 deg
test row 440
test row 440removed
pose_041754, middle finger, from the released 1,000 pose test split
lh_MFJ2 -0.0263 · lh_MFJ1 1.4917 · split 1.5180 rad · 87.0 deg · 220 of the 1,000 test poses are cut
test row 411
test row 411removed
pose_063847, first finger, also from the released test split
lh_FFJ2 1.4835 · lh_FFJ1 -0.0175 · split 1.5010 rad · 86.0 deg
train row 76309
train row 76309kept
pose_058090, little finger curling as one piece, kept
lh_LFJ2 0.8830 · lh_LFJ1 0.9613 · split 0.0783 rad · 4.5 deg · narrowest in the 11,000 read
train row 44637
train row 44637kept
pose_073444, ring finger curling as one piece, kept
lh_RFJ2 1.0128 · lh_RFJ1 0.9203 · split 0.0925 rad · 5.3 deg
Menlo robot-hand-poses-train · 11,000 of 100,000 poses read · Qwen2.5-VL 3B · 7 joint groups × 120 poses × 2 runs Wilson 95% · cut = poses where a tendon coupled pair is commanded over 1.20 rad apart · tendons and joint limits read from the Shadow Hand model in MuJoCo Menagerie