Results

PRIMARY ENDPOINT

S1: speed to a good design

A run's count is the number of its first design, counted from design 9, that seats and centres the part in at least 37 of 48 training grasps. Designs 1 to 8 are round 0 (the hand-designed entrants and seed designs), the same for both arms at each seed, so they do not count. A run that never gets there within 40 designs counts as 41. Fixed in the protocol before the prereg-v1 tag.

Best training success so far

Thin lines are runs, heavy lines each arm's median. A run whose round 0 already reached the target still needs one of its own designs (from design 9) to count.

Designs needed to reach the target

Exploratory, lift-only scoring (the part only has to come off the table): no difference between arms (p = 1.0). Both arms find a finger that lifts the part almost at once. The advantage is in seating and centring the part, which is what the next step in a work cell depends on.

SECONDARY ENDPOINTS (HOLM FAMILY)

Held-out success and search efficiency

Each run's frozen top-1 design was scored on suites neither arm saw during search (Nebius job aijob-e05nvfme46ss0q4wp4).

• H1: seated and centred, 100 unseen procedural conditions x 2 seeds

  Random (A)

  60.5%

  Nemotron (C)

  72.5%

  Difference

  +12.0 points

  p

  7e-05

  Holm-adjusted p

  0.00017

• H2: seated and centred, 64 real scanned objects

  Random (A)

  22.0%

  Nemotron (C)

  27.5%

  Difference

  +5.5 points

  p

  0.00077

  Holm-adjusted p

  0.00077

• S2: area under the best-so-far curve

  Random (A)

  0.703

  Nemotron (C)

  0.788

  Difference

  +0.085

  p

  5.5e-05

  Holm-adjusted p

  0.00017

All three reject the null at alpha 0.05 after Holm adjustment.

H2 by object width (planned split, reported outside the Holm family): objects 50 mm or narrower across the jaw (18 objects) 19.9% vs 24.1% (p = 0.012); wider than 50 mm (46 objects) 22.8% vs 28.8% (p = 0.001).

Real scanned objects are hard for every design in this space: the best arm seats them about a quarter of the time. The held-out real-object result says the Nemotron arm's picks generalise better, not that the fingers are ready for arbitrary objects.

Limits

• Everything is simulated, with rigid contacts. No finger was printed or tested on hardware for this result.
• One gripper geometry, six training parts, one task (seat, lift, hold).
• The effector tournament (hands, grippers, soft arm, suction, Nemotron-designed effectors) is product mode and not part of this pre-registered result.