Nemotron's first 4 proposals against the stock finger

Nemotron 3 Super read the results of 8 starting designs on 16 training grasps and proposed 4 new fingers; 3 of 4 passed the checks (2 model calls, 14,257 prompt and 3,473 completion tokens, $0.0074). The remaining slot failed the validator or the diversity check twice and was filled at random, as the protocol requires. Each lane is a design's own best grasp height on the training condition where their outcomes differ most: wire rod lying flat, radius 3 mm, length 120 mm, at (-30, 40) mm, yaw -60 deg, sitting (6, 6) mm and 15 deg away from where the controller aims.

PhysicsLocal run on a developer laptop, not a Nebius jobMuJoCo 3.15.0CPU executor1.60 s sim wall time14,445 steps
DesignHand-designed baseline, no model callHand-designed baseline entrant
Designnvidia/nemotron-3-super-120b-a12b via nebius-tokenfactory9,800 tokens11.7 s$0.0043round 1
DesignRandom sample (scrambled Sobol (validator-filtered))seed 17

t 00.00 / 00.00 s

Phase Settle

Settle
Descend

Baseline: stock finger

Loading replay

Wire rod lying flat, radius 3 mm, length 120 mm, at (-30, 40) mm, yaw -60 deg, sitting (6, 6) mm and 15 deg away from where the controller aims stock finger

Nemotron mutate 1

Loading replay

Wire rod lying flat, radius 3 mm, length 120 mm, at (-30, 40) mm, yaw -60 deg, sitting (6, 6) mm and 15 deg away from where the controller aims d09 finger

Random fallback (mutate 2)

Loading replay

Wire rod lying flat, radius 3 mm, length 120 mm, at (-30, 40) mm, yaw -60 deg, sitting (6, 6) mm and 15 deg away from where the controller aims d10 finger

Nemotron crossover

Loading replay

Wire rod lying flat, radius 3 mm, length 120 mm, at (-30, 40) mm, yaw -60 deg, sitting (6, 6) mm and 15 deg away from where the controller aims d11 finger

Nemotron explore

Loading replay

Wire rod lying flat, radius 3 mm, length 120 mm, at (-30, 40) mm, yaw -60 deg, sitting (6, 6) mm and 15 deg away from where the controller aims d12 finger

All lanes share one clock and one camera. The dashed square is the success height (object centre 100 mm above its resting height); the bar under each lane fills while the object stays above it and turns green at 2.0 s. The verdict appears when the scorer decides.

Scorer numbers

From the simulator's scorer, sampled every 10 ms.

Lane	Fingers	Outcome	Peak lift	Longest hold	Sim steps	Wall time
Baseline: stock fingerThe fingers never pinched the object before the lift started.	Stock	Missed grasp	0.0 mm	0.00 s	2,805	0.31 s
Nemotron mutate 1wire rod lying flat, radius 3 mm, length 120 mm, at (-30, 40) mm, yaw -60 deg, sitting (6, 6) mm and 15 deg away from where the controller aims	D09	Held 2.0 s	150.1 mm	2.01 s	3,015	0.27 s
Random fallback (mutate 2)The object was pinched but slid out before reaching the target height.	D10	Slipped during lift	51.9 mm	0.00 s	2,805	0.30 s
Nemotron crossoverThe fingers never pinched the object before the lift started.	D11	Missed grasp	0.0 mm	0.00 s	2,805	0.46 s
Nemotron explorewire rod lying flat, radius 3 mm, length 120 mm, at (-30, 40) mm, yaw -60 deg, sitting (6, 6) mm and 15 deg away from where the controller aims	D12	Held 2.0 s	150.0 mm	2.01 s	3,015	0.27 s

Task task_v1

Success: Centre rises 100 mm and stays 2.0 s, before 12 s
Grip force: 50 N
Max opening: 84 mm
Controller: Open loop, grasp-height setting k = 0.3

Phases

Settle0.0 s to 0.3 s
Descend0.3 s to 1.3 s
Close1.3 s to 2.6 s
Lift2.6 s to 4.1 s
Hold4.1 s to the end