Most data marketplaces are organized by content type or language. The robotics workflow needs an index by skill category. Each skill below points to the verticals where Anthrobase has captured it. Filter, scan, request a sample pack.
Reach, grasp, transport, place. The base manipulation primitive. Object variety from rigid containers through deformable produce, with practitioner-level fluency in handling speed and grip adaptation.
Two-handed tasks with role asymmetry, handoffs, and synchronization. The hardest gap in current humanoid training data. Weaving, kneading, threading, layering. India-dominant for one structural reason. Almost all skilled craft here is bimanual.
Force-dominant tasks. Insertion, alignment, joining, peg-in-hole at non-trivial tolerances. Force at the tool tip and the workpiece, sampled at 1 kHz across the contact event, frame-aligned with vision.
A tool changes what a hand can do. Practitioners with decades on a single tool perform manipulations no learned policy has seen. Catalogued by tool class, grip type, intended motion, and contact target.
Pouring, ladling, tempering, sprinkling, kneading. State that flows and shifts under the hand. The class of behaviour that breaks every rigid-body assumption in current robot learning stacks.
Sub-millimetre precision tasks. Embroidery, jewellery setting, hand stitching, beadwork. Practitioners with thirty plus years of muscle memory operating at a tolerance most robot policies cannot match.
Reorienting an object within a single hand using finger gating. The most dexterous primitive on the list. Coin work, tool flipping, fabric folding within one grip. Almost absent from public training data.
Tasks of one hour or longer, with stage boundaries, internal handoffs, tool switching, and ambient context. The horizon at which planning and recovery actually matter. Captured as continuous sessions, never spliced.