Tasks

Nine real-robot removal tasks in three families. Each has a preparation recipe you can repeat, held-out test conditions, and a success criterion you can measure. Each card opens a successful rollout next to a baseline failure mode.

    Protocol

    Rollouts end at success or at 120 s. Six core tasks carry the full input ablation; A5, B2 and C2 are extension scenarios.

    IDTaskPreparationHeld-outSuccess
    A1Barnacle model on a turtle shell0.35 g hot-melt adhesive, wait 60 sUnseen barnacle sizes, pose, colour; unseen shellDetached and lifted ≥ 10 mm for 1 s, no damage
    A23D-printed part on a build plate15 × 15 mm nano-adhesive pad, press 10 sTriangular and quad prisms; ± 30° yawDetached with clearance held, no violent ejection
    A3Blu-Tack on an acrylic plate≈ 20 g patch, press 5 N for 5 sBlue colour; ring and strip shapes; ± 50 mm placementResidue ≤ 5% by area
    A4Wood-ear model on a birch logHot-melt at the root, wait 60 sLog rotated ± 45°; specimen ± 30 mmDetached and lifted ≥ 10 mm for 1 s, no damage
    A5Blu-Tack residue on the fingertipAttach residue, randomise shape and poseClay; thin-sheet and thick-blob shapesFully removed by visual inspection
    B1Marker stain on a ceramic plateDraw stain in a 120 × 120 mm region; sponge wetted with 2 ml waterUnseen location, shape, colour≥ 90% of the region cleaned
    B2Ink mark on silicone skinDraw mark, wait 30 s; cotton tipUnseen location, shape, colour≥ 90% of the region cleaned
    C1Fossil model buried in sandBury at 15 mm, flatten the surfaceUnseen fossils; depths 10 and 20 mm; sand camouflage≥ 80% exposed, trench ≤ 20 mm deep
    C2Clump embedded in cat litterEmbed and level the litterColour-matched clump; ± 60 mm pose≥ 80% of the clump surface exposed

    Cleaned and exposed areas are measured by before/after pixel differencing in HSV space with a fixed threshold. "No damage" means no visible gouge or tear larger than 2 mm.

    Where to find the details. The task suite is how the hand and the force representation were evaluated. Objects, preparation steps, held-out conditions and all results are described in the paper. The open-source release covers the hand, the fingertip force sensor, the simulation models and the software.