Eigengrasp
- Also called: postural synergy, hand synergy
- Category: Planning & manipulation › Manipulation
- Tags: grasping, dexterous-hands, dimensionality-reduction, synergies, grasp-planning
A basis direction in a robot hand's joint space, derived from typical grasp postures, so that hand shapes can be described and searched with a few coefficients instead of every joint angle.
Definition
An eigengrasp is one of a small set of directions in a hand's joint space along which most useful grasp postures vary. A hand posture p is written as a mean posture plus a weighted sum of eigengrasps, p = p₀ + Σ aᵢeᵢ, so a hand with around twenty joints can be shaped with a few amplitudes aᵢ. The idea was introduced for robotic grasp planning by Ciocarlie and Allen, building on human-hand studies by Santello, Flanders and Soechting in 1998, which found that a few principal components of recorded grasp postures, called postural synergies, account for most of their variation. Eigengrasps are usually obtained by principal component analysis of such postures, either human data mapped onto the robot hand or postures designed for it.
The main use is grasp planning. GraspIt! includes an eigengrasp planner that searches jointly over the wrist pose and the eigengrasp amplitudes with simulated annealing, guided by an energy that rewards hand surfaces lying close to the object with suitable normals. With two eigengrasps, the search has eight dimensions, six for the wrist pose and two amplitudes, instead of more than twenty-five. Once a promising pre-grasp is found in the subspace, the fingers close in their full joint space until they make contact, and a grasp quality metric evaluates the result.
The low-dimensional subspace is also the main limitation. Postures that lie outside it, such as precise fingertip placements or unusual finger combinations, cannot be reached during the search, so grasps found this way cluster around a few typical hand shapes; recent dataset work notes that eigengrasp-planned grasps follow a narrow distribution and optimizes over the full joint space instead. Beyond planning, the same idea maps data-glove input onto robot hands in teleoperation, gives learning methods a reduced action space, and guides the design of underactuated hands in which one or two motors drive the fingers along fixed synergies. An underactuated hand builds such coupling into its mechanism, whereas eigengrasps are a choice made in software and can be dropped once a coarse grasp is found.
Why it matters
Dexterous hands have too many joints for naive search or exploration. Eigengrasps reduce that dimensionality using the structure of real grasps, and they explain both why early dexterous grasp planners worked and why their outputs lacked diversity, which matters when those outputs are used as training data.
Key applications
- Grasp planning for multi-fingered hands in GraspIt!
- Data-glove teleoperation mapped through a synergy subspace
- Reduced action spaces for learning dexterous policies
- Design and control of synergy-based underactuated hands
- Grasp synthesis for prosthetic hands
- Transferring grasp postures between hands with different kinematics
In practice
A researcher wants grasps for a five-fingered hand with over twenty joints on a set of household object models. In GraspIt! they load eigengrasps for the hand, defined by mapping principal components of human grasp postures onto the robot's joints, and run the planner, which varies the wrist pose and two amplitudes while scoring how closely the palm and fingers conform to each object. The best pre-grasps are then closed fully in simulation and ranked with the epsilon metric. Reviewing the results, the researcher notices that nearly every grasp is a variation of a power wrap and that pinches of small parts are missing, so they add a second stage that refines the best grasps in the full joint space to recover finer fingertip contacts.
Common misconceptions
- An eigengrasp is a grasp type, like a pinch or a power grasp — it is a basis direction in joint space; any single posture is a combination of several
- Eigengrasps can represent any grasp the hand can make — they span only a subspace, so precise or unusual postures fall outside it
- Eigengrasps must come from human data — they can also be designed by hand or computed from a robot's own successful grasps
- Using eigengrasps makes the hand underactuated — in planning they constrain only the search; a fully actuated hand can still move each joint independently, for example when closing on the object
Who should know this
- Dexterous manipulation researchers
- Robot hand and prosthetic hand designers
- Teleoperation engineers mapping human hands to robot hands
- Researchers generating grasp datasets
Related technologies
- GraspIt!
- Principal Component Analysis
- Postural Hand Synergies
- Pisa/IIT SoftHand
- Data Gloves
- Simulated Annealing
Related terms
- Grasp Planning
- Dexterous Manipulation
- Underactuated Hand
- Grasp Quality Metric
- Motion Retargeting
- Configuration Space