MARB · Robotic hand · September 19, 2026
Robotic hand leaderboard
A separate benchmark for assembling authored robotic-hand components. Amazing Hand and ORCA will be reported as distinct cases, not combined into one ranking.
Amazing Hand
Initial pilot complete · no ranked entries. The retained case uses 29 component designs in 226 positions. Three Astra attempts and three supervised Grok Bot attempts establish a baseline under the stated execution and scoring conditions. No complete reference-pose reconstruction was obtained; these observations do not establish a model ranking.
ORCA
In preparation · no ranked entries. Repeated import and rigid-placement checks passed for five flagged source components. Their known geometry defects remain documented exceptions, not successful validity checks. Source-version and parts-list conflicts remain unresolved.
How results will be compared
Each published comparison must identify the hand, kit and scoring versions, model and effort, runtime, assistance conditions, and attempted and graded run counts. Repeated trials must report failures and variation, not just the best export.
Hand results are not comparable with M3-CRETE scores. Initial preparation concerns digital assembly placement; it does not establish grasping performance, strength, safe motion or physical buildability. Berkeley Lite is deferred.
Interactive exhibit
Explore Halden Mk-IV: a 59-part hand concept with modeled servos and tendons, assembly animation and eight grip poses. This unscored visualization is separate from the Amazing Hand reconstruction results and does not establish physical buildability.
Next steps
Retain the Amazing Hand task and scoring for future model-update comparisons. ORCA still needs a matching authored assembly or accepted component placements before its baseline can proceed. The broader campaign remains incomplete; Berkeley Lite can be added later.