本文由 AI 分析生成
建立時間： 2026-07-02 來源： https://www.eng.yale.edu/grablab/pubs/Feix_THMS2016.pdf
Summary
This paper surveys 22 existing human grasp taxonomies from robotics, medicine, and biomechanics literature, extracts 211 grasp examples referenced across them, and synthesizes the largest non-redundant set into a single systematic taxonomy — 33 grasp types — organized along four axes: opposition type (Palm/Pad/Side), virtual finger assignment, power/intermediate/precision category, and thumb position (abducted/adducted). The 33 types can be further collapsed to 17 more general grasps if object shape/size is ignored.
本文比較文獻中 22 套既有人類抓握分類法,從中彙整出 211 個抓握範例,合成出目前最完整的單一分類系統——33 種抓握類型,依四個軸線(對向型態、虛擬手指分配、力量/中間/精準類型、拇指位置)系統化排列;若忽略物體形狀/大小,可再簡化為 17 種通用抓握類型。
Prerequisites
- Grasp definition (static, single-hand, orientation-independent) — the paper’s entire comparison hinges on a precise operational definition: “a static hand posture with which an object can be held securely with one hand, irrespective of hand orientation.” This excludes in-hand motion, bimanual grasps, and gravity-dependent grasps (e.g., hook grasp) — understanding this scope boundary is necessary to interpret which grasps made the final 33.
- Power vs. precision grasp distinction (Napier 1956, Landsmeer 1962) — foundational biomechanics concept: power grip = rigid hand-object coupling where the arm moves the object; precision handling = hand performs intrinsic movements without arm motion. The paper’s “intermediate” category is a later refinement of this binary.
- Opposition space and Virtual Finger (Iberall 1997) — the paper’s core organizing concepts. Opposition type (Pad/Palm/Side) describes the direction hands apply force; Virtual Finger describes how multiple real fingers merge into one functional unit opposing another unit. Both are prerequisite vocabulary for reading the taxonomy matrix (Fig. 4).
Core Idea
Rather than proposing grasp categories from first principles, the paper’s key contribution is a systematic literature synthesis method: register all grasps from 22 published taxonomies into one large comparison table (22 publications × 47 candidate grasp columns), merge grasps judged equivalent by hand configuration/object size/contact surface, exclude grasps that violate the static single-hand definition, and arrange the resulting union (33 types) along four largely orthogonal classification axes. The novel axis the authors add on top of prior work is thumb position (abducted vs. adducted) — because the thumb is the primary opposing digit, whether it can actively oppose the fingertips (abducted) or merely brace against finger sides / stay out of the way (adducted) turns out to meaningfully separate grasp types that share the same opposition type and VF assignment. The result is a taxonomy that is provably a superset of every reviewed prior taxonomy, and therefore the most complete available at the time.
Results
Not a benchmark paper — the “results” are the taxonomy itself plus statistics validating its coverage and structure, gathered from follow-up studies (Bullock et al.) that recorded real grasp usage by 2 housekeepers + 2 machinists (~4700 grasps observed):
| Metric | Value |
|---|---|
| Grasp examples found across 22 taxonomies | 211 |
| Candidate grasp types before filtering | 47 |
| Excluded (violate grasp definition) | 5 |
| Merged as minor variants | 8 |
| Final unique grasp types | 33 |
| Reduced set (merging within-cell grasps) | 17 |
| Grasp types actually observed in real-world usage study | 31 of 33 (all except Distal Type #19, Tripod Variation #21 — both highly specialized) |
| Coverage of real-world grasp duration/frequency using only the reduced 17-grasp set | 83.4% duration / 75.8% frequency |
| Correlation between a grasp’s literature-reference frequency and its real-world observed frequency | Pearson r = 0.36 (weak) — e.g. palmar pinch (#9) and tip pinch (#24) are over-cited relative to real use; lateral tripod (#16) is under-cited relative to real use |
Limitations
- Author-stated: Scope is restricted to static, stable, single-hand grasps — nonprehensile grasps (e.g., flipping a switch), in-hand manipulation, bimanual tasks, and dynamic grasping are explicitly excluded and left as “possibly best kept as separate classifications” for future work.
- Author-stated: Completeness cannot be proven from literature alone; the validation study (housekeepers + machinists) is limited to two work domains, so generalization to all environments/tasks is untested. Grasp-type frequency is shown to be strongly environment-dependent, which limits how far the frequency statistics (Fig. 5, Table I) generalize.
- Author-stated: Classification requires knowing both hand pose and hand-object contact type (e.g., palm-contact vs. fingertip-only distinguishes medium wrap #3 from prismatic-4-finger #6 despite similar hand shape) — this makes the taxonomy harder to apply for automated recognition systems that only sense joint angles without contact information.
- Unstated / noted during analysis: As a synthesis/survey rather than an empirical measurement paper, the taxonomy’s boundaries (which grasps count as “equivalent” across authors) rely on the authors’ subjective judgment when merging table entries — no formal inter-rater reliability is reported for the merging process itself (only for grasp confusion in the later usage study).
Reproducibility
- Code: not applicable (taxonomy/classification paper, no code artifact); taxonomy and supplementary comparison documents available at grasp.xief.net
- Datasets: The 33-grasp taxonomy itself is the artifact. Downstream validation used the Yale Human Grasping Dataset (Bullock et al. 2015) — video-recorded real-world grasps from housekeepers and machinists — but that dataset is a separate publication, not part of this paper.
- Compute: not applicable
Insights
This 2016 taxonomy is the standardization layer underneath a decade of downstream dexterous-grasping robotics work. It is directly consumed by Dexonomy Synthesizing All Dexterous Grasp Types in a Grasp Taxonomy (2025), which treats these 33 grasp types as the target label space for automatically synthesizing 9.5M robotic grasps — explicitly excluding the same two types (#19 Distal, #21 Tripod Variation) this paper found were unobserved in real-world usage, for the same reason (too object-category-specific: scissors, chopsticks). This is a clean example of a foundational HRI/biomechanics classification enabling a machine-learning application roughly a decade later — the “common terminology” goal stated in this paper’s abstract is exactly what let Dexonomy define its dataset’s label space without inventing a new taxonomy.
The finding that literature-citation frequency correlates weakly (r=0.36) with real-world usage frequency is also a broader methodological caution: grasp types that dominate robotics papers (e.g., precision pinches, because they’re easy to model analytically) are not necessarily the grasps humans use most — power grasps and the lateral grasp are under-represented in the literature relative to actual use.
Connections
- Dexonomy Synthesizing All Dexterous Grasp Types in a Grasp Taxonomy — directly consumes this taxonomy as its target grasp-type label space for automated grasp synthesis
- pomdar-dexterity-benchmark-anthropomorphic-robotic-hands — another dexterous-hand paper referencing the GRASP taxonomy
Raw Excerpt
“A grasp is every static hand posture with which an object can be held securely with one hand, irrespective of the hand orientation.” … Overall, 33 different grasp types are found and arranged into the GRASP taxonomy. Within the taxonomy, grasps are arranged according to 1) opposition type, 2) the virtual finger assignments, 3) type in terms of power, precision, or intermediate grasp, and 4) the position of the thumb.