ContactGrasp: Contact-Centric Grasp Synthesis

• ContactGrasp is a contact-centric approach that represents grasps via object-surface contact maps rather than traditional end-effector poses.
• It encompasses multiple frameworks, including multi-finger functional synthesis, dexterous pre-grasp datasets, and 6-DoF grasping in cluttered scenes.
• Evaluations show improved grasp selection and robust performance across diverse hand models using sample-and-rank optimization and contact-guided priors.

Searching arXiv for the specified ContactGrasp-related papers and closely related contact-centric grasping work. arXiv search query: ContactGrasp (Brahmbhatt et al., 2019, Xu et al., 28 Sep 2025) Contact-GraspNet (Sundermeyer et al., 2021) Contact2Grasp (Li et al., 2022) ContactGrasp is a contact-centric formulation of grasp synthesis in which object-surface contact, rather than only end-effector pose, serves as the primary representation for generating, transferring, or evaluating grasps. In the arXiv literature, the term designates more than one distinct system: a 2019 framework for functional multi-finger grasp synthesis from object-centric contact, and a 2025 robotic dexterous pre-grasp and grasp dataset integrated into GES-UniGrasp. Closely related work such as Contact-GraspNet extends the same contact-rooted principle to efficient 6-DoF grasp generation in cluttered scenes (Brahmbhatt et al., 2019, Xu et al., 28 Sep 2025, Sundermeyer et al., 2021).

1. Terminological scope and historical usage

The name “ContactGrasp” is not attached to a single canonical method across the literature. Instead, it marks a family of contact-driven formulations that differ in embodiment, supervision source, and optimization regime.

Usage	Paper	Core role
Functional multi-finger grasp synthesis from contact	"ContactGrasp: Functional Multi-finger Grasp Synthesis from Contact"	Object-centric contact maps guide sample-and-rank synthesis across hand models
Robotic dexterous pre-grasp and grasp dataset	"GES-UniGrasp: A Two-Stage Dexterous Grasping Strategy With Geometry-Based Expert Selection"	Contact-guided dataset and priors for RL-based dexterous grasp execution
Contact-centric 6-DoF grasp generation	"Contact-GraspNet: Efficient 6-DoF Grasp Generation in Cluttered Scenes"	End-to-end grasp generation rooted at observed point-cloud contacts

A recurrent misconception is that ContactGrasp denotes a single benchmark or a single optimization pipeline. The literature instead uses the term for at least two substantially different objects of study: first, a framework for reproducing functional contact patterns on anthropomorphic hands; second, a dataset-and-policy prior for dexterous robot manipulation. Related methods share the contact-centric premise but differ in whether they target human-hand synthesis, robot-hand retargeting, or parallel-jaw grasping (Brahmbhatt et al., 2019, Xu et al., 28 Sep 2025, Sundermeyer et al., 2021).

2. Functional grasp synthesis from object-centric contact

In its 2019 formulation, ContactGrasp addresses functional grasp synthesis for high-DoF anthropomorphic hands by taking as input object shape and an object-centric contact map that may be manually specified or obtained through demonstrations. A grasp is represented as the full hand configuration $\Phi$, including the 6-DoF palm pose and joint values for all articulated links. The central distinction is between stable grasping and functional grasping: a stable grasp resists disturbances, whereas a functional grasp additionally preserves affordances needed for post-grasp action, such as button access or unobstructed transfer surfaces (Brahmbhatt et al., 2019).

The contact representation is a multi-point object-centric map $c = \{(p_i,c_i)\}_{i=1}^N$, where $p_i$ are surface points and $c_i \in \{+1,-1\}$ denote attractive and repulsive samples. In the demonstration setting, $c_i$ is derived from thermal contact strength $t(p_i) \in [0,1]$ in ContactDB by thresholding at $\tau_t = 0.3$. This representation encodes both positive and negative evidence: attractive regions pull the synthesized hand toward demonstrated contact, while repulsive regions discourage contact in semantically important non-contact areas such as a wine-glass opening or a phone screen (Brahmbhatt et al., 2019, Brahmbhatt et al., 2019).

The synthesis pipeline is explicitly sample-and-rank. Diverse initial grasps are generated with GraspIt! using a coarse parameterization $\phi = (a,\theta,d)$, where $a$ is an approach point sampled uniformly on the object surface with approach direction $-n(a)$, $c = \{(p_i,c_i)\}_{i=1}^N$0 is a roll around the approach vector, and $c = \{(p_i,c_i)\}_{i=1}^N$1 is the offset distance from the surface. GraspIt!’s Simulated Annealing planner runs for 45K iterations per seed, and the top two full grasps $c = \{(p_i,c_i)\}_{i=1}^N$2 per seed are retained for refinement (Brahmbhatt et al., 2019).

Refinement minimizes

$c = \{(p_i,c_i)\}_{i=1}^N$3

where

$c = \{(p_i,c_i)\}_{i=1}^N$4

the repulsive activation is cylindrical,

$c = \{(p_i,c_i)\}_{i=1}^N$5

and the thumb term is

$c = \{(p_i,c_i)\}_{i=1}^N$6

The solver is DART with a modified Contact Prior and Levenberg–Marquardt optimization. Because contact is encoded on the object rather than on a specific hand, the same demonstration can be transferred to HumanHand, Allegro, and Barrett models without hand-specific retargeting (Brahmbhatt et al., 2019).

The method was evaluated on 19 household objects and two intents, “use” and “handoff.” For the HumanHand model, the median rank of the correct grasp was 3 for use and 1 for handoff, compared with 4484 and 3751 for GraspIt!, respectively. Average disagreement with human contact, measured by $c = \{(p_i,c_i)\}_{i=1}^N$7, was also lower than both GraspIt! and Human$c = \{(p_i,c_i)\}_{i=1}^N$8Allegro retargeting. These results operationalize ContactGrasp’s original claim: functionality can be encoded and transferred more reliably through contact patterns than through direct hand-pose retargeting (Brahmbhatt et al., 2019).

3. ContactGrasp as a robotic dexterous pre-grasp and grasp dataset

In the 2025 GES-UniGrasp system, ContactGrasp is a robotic dexterous pre-grasp and grasp dataset introduced to address a different limitation: existing reinforcement learning methods guided by grasp priors often induce unnatural behaviors in high-DoF robot hands, while large-scale procedural datasets can be compute-intensive and produce implausible contacts. ContactGrasp explicitly encodes task-relevant wrist orientation and thumb–index pinching coordination, and it is tightly integrated into a two-stage Geometry-based Expert Selection framework (Xu et al., 28 Sep 2025).

The dataset contains 773 object instances across 82 categories, with splits of 563 training objects, 132 test objects from seen categories, and 78 test objects from unseen categories. It is generated in simulation from object and table point clouds through a four-stage annotation pipeline. First, contact points are generated from the object point cloud $c = \{(p_i,c_i)\}_{i=1}^N$9 and table point cloud $p_i$0. The pipeline differentiates cylindrical from non-cylindrical objects, uses PCA on the tabletop to choose horizontal or vertical grasp strategy, performs full PCA alignment for non-cylindrical objects and $p_i$1-axis-restricted rotation for cylindrical objects, and refines five contact points through local optimization to ensure force closure and five-finger coordination. Second, the contact-derived grasp is retargeted to the robot hand through the topology descriptor

$p_i$2

with pronation–pitch coupling fixed at $p_i$3 and $p_i$4 (Xu et al., 28 Sep 2025).

Wrist and palm origins are defined as

$p_i$5

$p_i$6

and retargeting solves

$p_i$7

Third, collision-aware refinement is carried out in Isaac Gym with PD control

$p_i$8

and a contact-aware freeze rule

$p_i$9

Fourth, the final grasp pose $c_i \in \{+1,-1\}$0 is shifted by $c_i \in \{+1,-1\}$1 m along the approach direction and intermediate joints are reset to form the pre-grasp pose $c_i \in \{+1,-1\}$2 (Xu et al., 28 Sep 2025).

Within GES-UniGrasp, ContactGrasp supplies both pre-grasp and grasp priors. Policies start from $c_i \in \{+1,-1\}$3, aim to reach $c_i \in \{+1,-1\}$4, and then complete lifting or placement. The RL state includes joint states $c_i \in \{+1,-1\}$5, contact states, fingertip forces, fingertip positions in the wrist frame, wrist 6D pose $c_i \in \{+1,-1\}$6 and velocity, object position $c_i \in \{+1,-1\}$7, and pose difference to the target grasp. Reward shaping combines task terms with contact-consistency and smoothness terms:

$c_i \in \{+1,-1\}$8

where

$c_i \in \{+1,-1\}$9

This design targets unnatural finger extension, oscillation, and loss of directional consistency across fingers (Xu et al., 28 Sep 2025).

The execution layer uses geometry-based clustering and a gating network

$c_i$0

to select among experts. Objects are grouped by K-means and category-based clustering; expert training proceeds through intra-cluster central training, intra-cluster expansion, and inter-cluster generalization, with a third expert trained on hard failure cases. The gating backbone is CurveNet with 1024-point inputs, chosen over PointNet++ and Point-BERT for stability and generalization. The gating loss is

$c_i$1

Reported performance includes stage-wise success rates of 87.7% for contact point generation, 88.9% for retargeting, and 91.5% for collision-aware refinement; contact synthesis takes approximately 5 seconds per object. In policy execution, GES-UniGrasp achieves 99.4% success on train objects and 96.3% on both seen-category and unseen-category test sets (Xu et al., 28 Sep 2025).

4. Contact representation as an intermediate variable

ContactGrasp sits within a broader shift from direct pose regression toward explicit contact representations. The 2019 formulation encodes attractive and repulsive contact directly on the object surface and then optimizes hand configuration to reproduce that pattern. Several later systems generalize this principle by changing the granularity of the contact variable, the embodiment, or the conditional structure of the model (Brahmbhatt et al., 2019).

One major branch reduces grasp generation to contact-conditioned pose synthesis. Contact-GraspNet roots a full 6-DoF parallel-jaw grasp in an observed point-cloud contact and predicts only local orientation and width, thereby reducing the problem to a 4-DoF contact-centric representation with

$c_i$2

This formulation is explicitly designed for cluttered scenes and avoids direct prediction in unconstrained $c_i$3 (Sundermeyer et al., 2021).

A second branch treats contact as a latent intermediate variable. Contact2Grasp factorizes the mapping as $c_i$4, approximating

$c_i$5

with a ContactCVAE for diverse contact-map generation and a second network that maps contacts to MANO hand poses. This suggests a strong structural affinity to ContactGrasp’s original object-centric premise, but with learned stochastic contact generation rather than demonstration-derived maps (Li et al., 2022).

A third branch increases contact resolution or controllability. GrainGrasp predicts separate contact maps for thumb, index, middle, ring, and pinky fingertips, then optimizes a MANO hand using contact-alignment and directional-consistency energies. RegionGrasp conditions grasp generation on an explicitly designated object region and requires the thumb fingertip to touch that region, operationalizing contact controllability through a binary region mask and a CVAE. CMG-Net, in turn, formulates BarrettHand grasp synthesis around fingertip contacts and reduces the grasp parameterization to

$c_i$6

mapping contact geometry analytically back to hand pose and joint configuration (Zhao et al., 2024, Wang et al., 2024, Wei et al., 2023).

Taken together, these systems show that ContactGrasp is less a single algorithm than a design pattern: object-centric or scene-centric contact serves as the shared latent interface between geometry, kinematics, and task constraints. The specific contact variable may be a binary attractive/repulsive map, a five-point fingertip set, a region mask, a fingertip-wise probability field, or a contact-rooted local grasp frame, but the computational role remains the same.

5. Evaluation regimes and reported performance

Evaluation in the ContactGrasp literature is method-specific rather than standardized. Functional human-hand synthesis, robot-hand dexterous execution, and parallel-jaw picking in clutter use different success definitions, different embodiments, and different stability proxies. Direct numerical comparison across these settings is therefore inappropriate, even when all methods are contact-centric (Brahmbhatt et al., 2019, Xu et al., 28 Sep 2025, Sundermeyer et al., 2021).

System	Evaluation regime	Reported outcome
2019 ContactGrasp	Rank of correct functional grasp and disagreement with demonstrated contact across 19 objects, three hand models, two intents	HumanHand median rank 3 for use and 1 for handoff; lower $c_i$7 than GraspIt! and Human$c_i$8Allegro retargeting
2025 ContactGrasp within GES-UniGrasp	Stage-wise synthesis success and RL execution success on 773 objects in 82 categories	87.7% contact generation, 88.9% retargeting, 91.5% refinement; 99.4% train and 96.3% test success
Contact-GraspNet	Real robot picking of 51 unseen objects in 9 cluttered scenes	90.20% overall success, 84.31% first-attempt success, 0.19–0.28 s inference

The 2019 framework evaluates whether contact-driven optimization can recover functional intent. Its principal signals are the rank of the correct grasp under residual scoring and the contact disagreement term $c_i$9. The 2025 dataset evaluates two separate processes: contact-guided synthesis itself and the downstream RL policy’s ability to lift to 0.2 m above the table and maintain the object for 200 simulation timesteps. Contact-GraspNet evaluates robotic clutter picking under first-attempt and overall success, with inference speed explicitly included because the system is designed for closed-loop use (Brahmbhatt et al., 2019, Xu et al., 28 Sep 2025, Sundermeyer et al., 2021).

This heterogeneity clarifies an important point: “contact-based” does not imply a shared empirical protocol. In some papers contact is judged by functional agreement with demonstrations; in others it is judged by policy convergence, lifting stability, or physical pick success in clutter.

6. Limitations, misconceptions, and prospective directions

A persistent misconception is that ContactGrasp methods necessarily optimize full contact physics. The literature is more heterogeneous. The 2019 framework does not explicitly include a force-closure stability term inside $t(p_i) \in [0,1]$0, relying instead on initialization and intersection avoidance. The 2025 ContactGrasp dataset states that contact points are refined “to ensure force closure and five-finger coordination,” yet explicit friction-cone inequalities and grasp wrench space formulations are not provided. Contact-GraspNet removes colliding simulated grasps from training data and performs collision checks at inference, but it does not prescribe analytic force-closure checks during learning (Brahmbhatt et al., 2019, Xu et al., 28 Sep 2025, Sundermeyer et al., 2021).

The concrete limitations are correspondingly diverse. The 2019 method is local and sensitive to initialization; some hand models cannot realize intent-specific contact because of morphology. The 2025 dataset relies on handcrafted heuristics for contact synthesis, reports that the gating network may have limited generalization, and has not yet been deployed on physical hardware; detailed file formats, URLs, licensing terms, object material statistics, and hardware specifics are unspecified. Contact-GraspNet reports lower confidence on thick objects near maximum width, reduced confidence on very small objects, and sensitivity to segmentation errors, even though it remains robust to under-segmentation (Brahmbhatt et al., 2019, Xu et al., 28 Sep 2025, Sundermeyer et al., 2021).

A second misconception is that contact-centric representations are inherently anthropomorphic. That is true for some systems but not for all. The 2019 ContactGrasp inherits human contact from ContactDB and explicitly targets functional grasps across human-like and robot hands. The 2025 ContactGrasp dataset explicitly encodes wrist orientation and thumb–index pinching coordination to produce more human-like robot grasp priors. By contrast, Contact-GraspNet is class-agnostic and parallel-jaw; its contact-centric design is geometric rather than anthropomorphic (Brahmbhatt et al., 2019, Xu et al., 28 Sep 2025, Sundermeyer et al., 2021).

The current trajectory of the field suggests three concrete extensions. First, semantic selection can be layered onto contact maps, as in "SECOND-Grasp: Semantic Contact-guided Dexterous Grasping," where vision-language reasoning proposes semantically meaningful contact parts before geometric refinement (Shin et al., 13 May 2026). Second, explicit force modeling can be coupled to contact generation, as in "Force-Aware 3D Contact Modeling for Stable Grasp Generation," where normal force is discretized into levels and used in an acceleration-minimization stability formulation (Chen et al., 17 Nov 2025). Third, region-level robustness can be formalized through independent contact regions rather than single contact points, as in "Fast Estimation of Globally Optimal Independent Contact Regions for Robust Grasping and Manipulation" (King et al., 10 Jun 2025). This suggests that future ContactGrasp-style systems may increasingly combine contact priors, semantic constraints, and force-aware stability criteria rather than treating them as separate stages.