Which effector for the L-bracket?

The finalists, side by side

One condition for every lane: 3 mm off centre across (train_v2/l_bracket/c2, seed 7936365254780240449), on one clock. Each lane re-ran its scored screen episode with recording on and reproduced it bit for bit. Bodies are drawn from the collision shapes the simulator used.

Families

Ranking (registered)

Eligible: certified, the nominal grasp passes, no simulator reset and no penetration over 1 mm in the nominal test. Ranked by certificate score, then the smaller placement tolerance, then screen success; within 0.05 and the same tolerance the simpler effector wins (fewer actuators, then lower mass, then higher fidelity).

Certificates

The pre-registered battery, unchanged: the nominal grasp, the part 1, 2 and 3 mm off where the wrist aims on both axes, a 1 g shake on three axes and a 0.5 N s push at the part's top on two. Green: registered; amber dash: held but not registered; red: dropped. Jitter is off.

Budget and fairness

Same for every candidate

yes (wrist servos and force limits, solver, the part's contact parameters, phase timings, conditions and seeds, read from each compiled model)

Conditions

c0: +0 mm (seeds 1213963221845979623, 1593426842636724785)c1: +2 mm (seeds 6343790114632006567, 97196857171235771)c2: -3 mm (seeds 7936365254780240449, 727452150277789134)c3: +1 mm, 6 deg (seeds 7111270448186729677, 2495793148784487793)

Physics

276 screen and reference episodes, 216 certificate episodes; 0 simulator resets, 1 over 1 mm penetration

Signature

95be2891a5bcc59c1ef46168ef55ed92c0d8e975e2ca3d2c2d36f2caa3729253

Caveats

• Every family: a top-down 4-DoF wrist (no roll or pitch), open-loop timing, no perception (the placement error is unknown to the controller), soft contacts and a 1 ms timestep. The recommendation is a simulation ranking under these conditions, not a hardware guarantee. The jaw's registration reference is its centre line; every other family's is repeatability against its own nominal episode, and each family runs at its own declared actuation limits.
• Industrial gripper: Rigid models with the upstream actuator limits; the adaptive grippers are our own geometry. Scripted closing, no tactile feedback.
• Multi-finger hand: Simplified rigid models with no skin compliance, driven by scripted open-to-closed synergies: no tactile feedback, no regrasp. Actuator limits are the upstream model's. Rubber pads would grip better than the part's mu 0.3 used here, so results are conservative for padded hands.
• Designed morphology: Rigid models generated from a recipe and driven by scripted synergies. The random control from the same space is shown beside the Nemotron designs; no superiority is claimed.
• Parallel jaw: The core experiment's own scene, checks and scorer. Registration is measured against the jaw's centre line.
• Soft arm: Approximation: a chain of rigid capsules with spring-damper joints, curled by a tension-limited cable. Not modelled: continuum mechanics, material non-linearity, suckers. Verify on hardware.
• Suction cup: Approximation: the seal is a geometric test gating an adhesion actuator capped at vacuum times lip area. Not modelled: air flow, leaks, porosity, vacuum build-up time, lip hysteresis. Treat as a screen, then verify on hardware.

MuJoCo 3.15.0 · code 790c47d65853 · task 8168db801332 · run effectors-l-bracket · 2026-10-17T11:54:22Z