EAGG Embodiment-Aligned Grasp Generation

arXiv 2606.18092

EAGG

Embodiment-Aligned Grasp Generation via Geometry-Aware Graph Conditioning

Wanhao Niu, Qiyan Ke, Yuan Sun, Hao Sun, Jie Xu, Muyuan Ma, Ruiqi Hu, and Fuchun Sun

Overview Video

Cross-object and cross-end-effector grasp generation.

EAGG generates object-conditioned grasps for heterogeneous robotic hands and grippers by conditioning the generator on both object geometry and end-effector structure.

Core Idea

Align embodiment structure instead of suppressing it.

Different hands and grippers have different topology, joint spaces, closure behavior, and contact geometry. EAGG keeps those differences explicit and turns them into geometry-aware conditioning for a shared grasp generator.

Embodiment-specific control basis

Each end effector keeps a low-dimensional control space without forcing all hands into one raw joint parameterization.

Topology-aware graph

The end-effector graph preserves kinematic organization, coupling, and morphology-specific structure.

End-effector cognition

A frozen backbone extracts geometry-aware tokens from the current articulated hand or gripper state.

Iterative geometry injection

Conditioning is refreshed during sampling so the generator remains synchronized with evolving grasp geometry.

Method

Geometry-aware graph conditioning for heterogeneous end effectors.

EAGG combines object point tokens, grasp tokens, and end-effector tokens inside a transformer generator. The hand/gripper tokens are produced from topology-aware graph features and a canonical end-effector cloud, then updated through iterative geometry injection.

EAGG method overview

Selected Results

Ranked visualizations and compact quantitative summaries.

Checkpoint inference is followed by ranked candidate selection and mesh rendering for bundled clean objects and released end effectors. The tables below summarize the paper's representative simulation, cross-regime, and hardware results.

Allegro generalization across objects

Cross-object generalization

Allegro is applied to multiple clean demo objects, with selected generated grasps rendered as object-hand meshes.

Mug generalization across end effectors

Cross-end-effector generalization

The mug object is fixed while the generated grasp is rendered for different hands and grippers.

Representative Simulation Comparison

End Effector Method SR (%) CD (cm) PD (cm) CC RGR (%)
WSG-50 EAGG Unified 14.50 0.40 0.77 1.262 8.76
WSG-50 EAGG Specialized 16.22 0.33 0.64 1.817 1.73
Robotiq 3F EAGG Unified 78.42 1.14 1.48 2.512 1.20
Robotiq 3F EAGG Specialized 79.26 0.78 1.07 2.615 0.18
Allegro EAGG Unified 86.46 0.50 1.04 2.564 0.12
Allegro EAGG Specialized 87.59 0.41 0.91 2.685 0.00

Cross-Regime Simulation Summary

Regime End Effectors SR (%) CD (cm) PD (cm) CC RGR (%)
Training Unified 6 avg. 56.17 0.56 1.04 2.081 3.16
Training Specialized 6 avg. 57.27 0.40 0.79 2.254 1.21
Held-out Adapted 2 avg. 36.68 0.36 0.62 2.196 3.63
Zero-shot Adapted 2 avg. 33.25 0.56 0.67 2.332 0.00

Real-World Hardware Evaluation

Setup Object Groups Objects Attempts Average SR (%)
UR5 + FreedomHand A-E 57 62 91.94
UR5 + DaHuan AG95 A-E 57 60 95.00
SOARM101 A-E 57 64 89.06

SR denotes success rate, CD contact distance, PD penetration depth, CC contact count, and RGR repeated grasp rate.