GRASP Taxonomy: Human & Robotic Grasp Types

• GRASP Taxonomy is a structured classification defining 33 canonical grasp types in power, precision, and intermediate categories based on extensive human and robotic studies.
• It integrates semantic, functional, and geometric aspects to represent contact patterns and force directions for reproducible hand-object interactions.
• Recent applications in grasp synthesis, vision-language modeling, and dynamic motion generation highlight its pivotal role in advancing dexterous manipulation research.

The GRASP taxonomy is a comprehensive, hierarchically structured classification of human and robotic hand grasp types, established to standardize the representation, analysis, and synthesis of dexterous manipulation. It originates from systematic studies of human prehension (notably Feix et al., 2015), formalizing 33 canonical grasp types across three principal categories—Power, Precision, and Intermediate. The taxonomy encodes grasp intent not only by postural anatomy and contact geometry but also by semantic, functional, and affordance-based properties, enabling reproducible datasets, scalable generative models, and physically grounded simulation across robotic and biomechanical applications. Recent works systematically integrate the GRASP taxonomy as an explicit condition in dataset generation, vision-language modeling, type-aware grasp synthesis, and motion generation, ensuring that the space of possible grasps is both exhaustive and semantically interpretable (Chen et al., 26 Apr 2025, Zhang et al., 3 Dec 2025, Shi et al., 21 Sep 2025, Zhang et al., 3 Dec 2025, Borràs et al., 2019, Augenstein et al., 25 Sep 2025).

1. Historical Development and Structure

The GRASP taxonomy, initially formalized by Feix et al. (2015), distills decades of neuroscience, biomechanics, and robotics research into a unified framework. The taxonomy’s structuring principle is the opposition geometry—defining whether and how the palm, fingers, or fingertips interact with the object. This foundation explicitly distinguishes:

• Power Grasps: Full-hand contacts with the palm contributing major force transmission (e.g., Cylindrical, Large Diameter).
• Precision Grasps: Force applied primarily by the distal phalanges (fingertips), relying on well-localized contacts (e.g., Tip Pinch, Tripod, Lateral Pinch).
• Intermediate Grasps: Both palm and fingertips contribute, with hybrid contact patterns (e.g., Spherical, Key, Modified Tripod) (Chen et al., 26 Apr 2025, Zhang et al., 3 Dec 2025).

Formally, the taxonomy is often represented as a two-level or tree-structured hierarchy: a coarse category ($T_{\rm coarse}$) {Power, Precision, Intermediate} and a fine category ($T_{\rm fine}$), where each grasp type $t_g^i$ is a child of a coarse class (Zhang et al., 3 Dec 2025, Augenstein et al., 25 Sep 2025). At the leaf level, each of the 33 types is characterized by a tuple:

• Opposition type (pad/palm/side)
• Active fingers ($S_f$), contacting links ($S_l^k$ per finger k)
• Contact region pattern ($S_c$)
• Force direction ($v$) and grasp center ($c$)

This structure allows unambiguous semantic and geometric reasoning about any grasp in the taxonomy.

Level	Notation	Example
Coarse	$T_{coarse}$	Power, Precision, Intermediate
Fine-grained	$T_{fine}$	Tip Pinch, Hook, Tripod

2. Taxonomy Integration in Robot Grasp Synthesis and Learning

Recent grasp-synthesis frameworks implement the GRASP taxonomy as a generative prior or as explicit annotation in large-scale datasets. For instance, Dexonomy (Chen et al., 26 Apr 2025) synthesizes 9.5 million grasps over 10,700 objects, strictly covering 31 of the 33 Feix taxonomy types (two excluded due to human specificity). The process begins with human-annotated type-specific grasp templates (joint angles, contact points/normals), followed by global object alignment, contact-aware refinement, and physically grounded force-closure validation. The taxonomy guarantees type diversity and reproducible mapping from intent to plausible realizations.

Each grasp type $G_i$ is stored as a template; valid hand-object arrangements are synthesized according to

$$L_{\text{align}} = k_p \sum_{i=1}^m \| p^h_i - p^o_i\|^2 + k_n \sum_{i=1}^m \| n^h_i - n^o_i\|^2$$

where template contacts $(p^h_i, n^h_i)$ are matched to object contacts $(p^o_i, n^o_i)$ (Chen et al., 26 Apr 2025).

Force-closure is verified via a convex program (LCQP), ensuring that grasps for each type truly adhere to Feix’s physical intent.

The taxonomy, thus, enables not only exhaustive sampling but also skill transfer between different robotic hands and object classes, provided that the contact and postural constraints for each type are met.

3. Grasp Taxonomy in Vision-Language and Cognitive Models

Integrating language and vision, OmniDexVLG (Zhang et al., 3 Dec 2025) and HOGraspFlow (Shi et al., 21 Sep 2025) extend the GRASP taxonomy into multimodal semantically controlled generative models. Explicit tokens encoding grasp type, contact semantics, and functional affordances are incorporated into both training and inference:

• OmniDexVLG uses language tokens such as “grasp type = tripod”, which, via a vision-LLM (VLM), produces a conditioning vector $\mathbf c$ sculpting the downstream denoising process to respect the taxonomy (Zhang et al., 3 Dec 2025).
• HOGraspFlow maintains a learnable taxonomy codebook, where a soft mixture $\hat \gamma_j = \sum_{k=1}^K \pi_{j,k} \gamma_k$ (with $\pi$ the predicted grasp-type distribution) directly conditions the diffusion network, ensuring that synthesized poses are categorically faithful to the human taxonomy (Shi et al., 21 Sep 2025).

Both frameworks rely on datasets (e.g., HOGraspNet) where each frame is labeled with a categorical type from the Feix taxonomy, grounding synthesis and recognition in the same structured vocabulary.

4. Quantitative Embedding and Dynamics: Taxonomy-Aware Latent Spaces

The taxonomy is not merely categorical; its internal structure shapes continuous latent spaces for motion and posture generation. Taxonomy-aware dynamical models embed the hierarchical relations between grasp types into the geometry of the latent representation.

In Taxonomy-aware Dynamic Motion Generation (Augenstein et al., 25 Sep 2025), grasp types are treated as nodes in a rooted tree (root → {power, precision} → subclasses → leaf postures). Trajectories between grasp types are modeled in Lorentzian hyperbolic space ($H^D$), enforcing that the geodesic distance $d_{H^D}(x_i, x_j)$ between two latent points $x_i, x_j$ reflect their graph distance $d_G(c_i, c_j)$ in the taxonomy:

$$\varphi(X) = \sum_{i < j} \left( d_G(c_i, c_j) - d_{H}(x_i, x_j) \right)^2$$

This “stress” penalty, jointly with GP dynamics, ensures that the latent representations preserve inter-type taxonomic relationships. The hyperbolic manifold is particularly well-suited for hierarchical trees, capturing exponential branching widths at each subtree.

5. Semantic Reasoning, Affordances, and Functional Integration

The GRASP taxonomy extends beyond morphology to encode opposition type, functional affordance, and semantic intent. In OmniDexVLG, each fine-level grasp type $t_g$ is associated with a configuration template $G = \{ S_f, S_l, S_c, v, c \}$, capturing:

• The set of contacting fingers and links
• The spatial contact pattern (e.g., antipodal pinch, three-point tripods)
• The force direction and intended object use-case

This representation enables retrieval-augmented reasoning and functional matching. For example, a “Hook” grasp is described by four fingers curled around a handle (palm-opposition), with the opposition normal aligned to the handle arc, and is selected for tasks such as suitcase lifting or bowl orbiting (Zhang et al., 3 Dec 2025).

Grasp Type	Category	Opposition	Fingers	Use-Case
Two-Finger Pinch	Precision	Pad	T, I	Small cylinder, pen
Tripod	Precision	Pad	T, I, M	Spoon, bowl rotation
Hook	Power	Palm	I, M, R, Li	Suitcase, handle lift

Dataset and simulation pipelines sample grasps based on affordance maps, opposition templates, and taxonomy-aware force-closure validation, ensuring both semantic and physical coherence.

6. Extensions: Multi-Object and Deformable Object Taxonomies

While the GRASP taxonomy focuses on single-object prehension, recent extensions address multi-object and non-rigid manipulation. The MOG taxonomy (Sun et al., 2022) organizes 12 multi-object grasp types into shape-based (e.g., cup, funnel) and function-based (e.g., scissors, tripod) categories, using similar opposition and closure concepts but framed for simultaneous handling of object sets.

In deformable contexts, such as cloth manipulation, “A Grasping‐centered Analysis for Cloth Manipulation” (Borràs et al., 2019) demonstrates that classical taxonomies fail, proposing instead a virtual finger (VF) framework. Here, the grasp is a tuple of geometric virtual fingers $\{VF_i, VF_j\}$ (points, lines, planes, both intrinsic and extrinsic), with “prehensile” grasp if dimensions match, and “non-prehensile” otherwise:

$$f(\{vf_i, vf_j\}) = \begin{cases} \text{prehensile}, & \dim(vf_i) = \dim(vf_j)\ \text{non‐prehensile}, & \text{otherwise} \end{cases}$$

This abstraction provides a task-centric vocabulary for benchmarking and gripper design even in highly deformable scenarios.

7. Practical Impact and Benchmarking

Adoption of the GRASP taxonomy across grasp dataset creation (Chen et al., 26 Apr 2025), generative modeling (Zhang et al., 3 Dec 2025, Shi et al., 21 Sep 2025), and dynamic motion synthesis (Augenstein et al., 25 Sep 2025) has established a reproducible standard for dexterous manipulation research. Empirically, taxonomic coverage is directly linked to increased grasp diversity, task transfer, and real-world success (e.g., Dexonomy: 9.5M grasps, 31 types, 82.3% real-world type-conditional success rate (Chen et al., 26 Apr 2025)).

Taxonomy-driven pipelines also guide future gripper design, benchmark complexity scaling, and evaluation: the ability to realize grasps across all taxonomic types is now a de facto benchmark for generalist manipulation platforms. The taxonomy’s abstraction over hand shape and semantics ensures extensibility to emerging domains such as multi-object, cloth, and affordance-based robotic manipulation.

---

In summary, the GRASP taxonomy provides a rigorous, extensible, and semantically rich foundation for describing, synthesizing, and evaluating dexterous hand-object interaction, permeating state-of-the-art methods in dataset generation, generative modeling, language-guided manipulation, and dynamic synthesis across a broad landscape of robotic and human manipulation research (Chen et al., 26 Apr 2025, Zhang et al., 3 Dec 2025, Shi et al., 21 Sep 2025, Augenstein et al., 25 Sep 2025, Borràs et al., 2019).