What is Dexterous Hand Data?

A dexterous hand moves its fingers independently (the Allegro Hand has four fingers and 16 joints, and the LEAP Hand is a lower-cost research alternative), so its data records every joint, not an open-or-closed state. The tasks differ in kind as well: rotating an object inside the hand, regrasping, thumb opposition. None of these reduce to a grasp pose, which is part of what makes the data expensive.

Capture is usually teleoperation through a glove or a retargeted human hand. Retargeting is the hard step: a human hand has more degrees of freedom than most robot hands, so the mapping has to choose what to preserve (fingertip positions, joint angles or contact), and different choices produce visibly different demonstrations of the same human motion.

Hands fitted with tactile sensing add a contact stream that has to stay calibrated, and a good sample names both the hand and the retargeting convention, because a trajectory without them cannot be reproduced. Dexterous collection is quoted at a multiple of simple gripper collection, and it is the category most often commissioned to order rather than bought off the shelf.

Types of in-hand skills

In-hand manipulation is the family a two-finger gripper cannot do at all. Rotating an object within the palm, whether it pivots on the fingertips or slides against them, is the classic case. Regrasping is the second: releasing a hold and taking a new one, which is how a hand reorients an object it cannot turn directly. Both are captured across every joint, so the episodes carry more state than a gripper recording ever does.

Functional tasks are the second group, closer to ordinary manipulation: precision pinching, thumb opposition, insertion with fingertip contact, tool handling. What makes them dexterous is that the fingers do work a wrist used to do. A third group exists mainly for sensing: tasks whose purpose is to feel an object's shape, weight or surface, where the contact record is the point rather than the motion.

The data also varies by how it was captured. Glove-based teleoperation records the operator's hand directly. Camera-based hand tracking does the same without the glove, at lower fidelity. Retargeting from full-body motion capture is the most indirect source and needs the most processing, because mapping a human hand onto a robot hand is a choice, not a conversion.

What to specify when buying

The hand model comes first, named exactly. Joint count, finger arrangement and whether the hand carries tactile sensing all change what the trajectories mean, and data collected on one hand rarely transfers to another. The retargeting convention is the second specification: fingertip positions, joint angles or contact. The same human motion mapped two ways produces two different datasets.

The object list matters more here than in gripper work. Dexterous tasks are usually object-specific (a small cube, a cylinder, a tool with a handle), and size, weight and compliance decide whether the skill is even possible on the target hand. A brief naming the objects and the skill family (rotate, regrasp, pinch) is more useful than one asking for dexterous manipulation in general.

Dexterous collection is priced well above gripper collection, and most of it is commissioned to order rather than bought off the shelf, because the hardware is uncommon and the skills are slow to record. If the project includes tactile sensing, say whether the sensor stream has to stay calibrated across sessions; that requirement changes the collection schedule, not just the price.