Method and limitations

The gripper descends, closes with a fixed force, lifts and holds. A trial succeeds when the object's centre rises at least 100 mm above where it rested and stays there for a continuous 2.0 s, finishing before 12 s. Every failure gets exactly one label:

Missed grasp

The fingers never pinched the object before the lift started.

Slipped during lift

The object was pinched but slid out before reaching the target height.

Pushed out on close

Closing the fingers knocked the object away.

Dropped during hold

The object reached the target height but did not stay there for the full hold.

Hit the table

A finger or the palm pressed into the table.

Held off its seat

The part was lifted and held, but not seated, centred and upright in the jaws.

Physics reset

The simulator reset its state before the outcome was decided, so the episode cannot count as a hold.

Physics unreliable: overlap over 1 mm

The jaws and the part passed into each other by more than 1 mm while it was held, so the hold cannot be trusted.

Jaws not buildable

The recipe's jaws could not be built for this part, so it scores as a failure without a simulation.

Physics checks run alongside: a finger pressing into the table, deep interpenetration, or an object moving impossibly fast all fail the trial, so a design cannot win by exploiting the simulator.

The simulated gripper matches an open-source parallel gripper: 84 mm maximum opening and 50 N of grip force. Every design uses the same gripper, the same open-loop controller and the same force; only the fingers change. Fingers are modelled in PLA and built from 14 boxes each, and the printable mesh is the union of exactly those boxes.

A design is ten numbers. The model proposes them through a forced tool call, and a validator rejects anything unprintable or unsafe before it reaches the simulator.

Length

40 to 120 mm

Thickness

4 to 16 mm

Width

10 to 30 mm

Bend angle

-15 to 45 deg

Bend position

0.2 to 0.9

Contact face

Flat, Concave arc, V groove

Face recess

0 to 8 mm

Tip profile

Blunt, Rounded, Chamfer, Hook

Tip feature

0 to 15 mm

Taper

0.5 to 1

Nemotron with measured feedback is compared with random search under an equal budget of simulated designs per run, standing in for print-and-test cycles. Tokens, dollars and wall time are reported separately.

The primary endpoint, S1, is speed to a good design: how many simulated designs a run needs before its best training success reaches a pre-registered threshold, with runs that never reach it censored at the budget. Held-out success (H1) is secondary. CPU MuJoCo is the executor of record. The pre-registered runs are set to use Nebius CPU instances; the pilot ran the same CPU executor on a Nebius GPU machine type, as the compute page shows.

Each run freezes its best design before any held-out trial. The held-out suite uses object sizes and poses the search never saw, plus real household objects from Google Scanned Objects (CC BY 4.0). The hypothesis, budget, number of runs, seeds and analysis script are written down and tagged before the main runs start, and the result is published whether it is a win, a tie or a loss.

The results page shows the run-level estimates; the compute page lists the jobs behind them.

• Nebius Token Factory serves NVIDIA Nemotron 3 Super, which proposes the designs. Every request and response is cached by hash, so the search can be replayed without API calls.
• Nebius Serverless Jobs run the scored MuJoCo trials from one pinned container image, writing results to Nebius Object Storage.
• This site is static. Replays are the simulator's own body poses at 30 frames per second, drawn in your browser.

• Everything here is simulated. No printed finger has been tested on hardware yet.
• Fingers are rigid and approximated by boxes; printed PLA bends a little under load.
• The controller is open loop: it does not sense contact or correct a slipping grasp.
• One friction coefficient governs every object contact; real finger pads and objects vary.
• Bouts recorded during development ran on a laptop, not on Nebius. Each bout's provenance chips say where it ran; only Nebius runs feed the results.
• Replays store body poses only, so contact forces are not drawn.