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

A unified generator for object-conditioned grasp synthesis across heterogeneous robotic hands and grippers, aligned through embodiment topology and articulated geometry.

Paper Code Checkpoints Watch Video
EAGG cross-end-effector grasp generation overview
Multi-object inputs and multi-hand inputs converge into EAGG, then branch into cross-object and cross-hand grasp generation.
Unified Generator One model operates across heterogeneous end effectors without a shared raw joint parameterization.
Embodiment Graph Topology-aware conditioning preserves kinematic structure and morphology-specific coupling.
Geometry Injection End-effector tokens are refreshed during sampling as articulated geometry evolves.
Hardware Transfer Evaluated in simulation and real-world setups across hands, grippers, and object groups.

Demo

Watch EAGG move across embodiments.

The video summarizes generated grasps, cross-end-effector behavior, and the visual result of geometry-aware conditioning.

Abstract

Align embodiment structure instead of hiding it.

EAGG treats embodiment as a geometric and topological signal, not a static label.

Cross-end-effector grasp generation seeks a single model that can synthesize grasps for heterogeneous end effectors while preserving generalization across novel objects. Existing generators are typically designed for fixed embodiments or encode embodiment identity with static labels, weakening transfer when topology, actuation coupling, and contact geometry differ.

EAGG addresses this with an embodiment-aligned representation: each end effector keeps a topology-aware graph and a PCA-based low-dimensional control space. A frozen end-effector-cognition backbone converts articulated states into geometry-aware tokens, while iterative geometry injection keeps conditioning synchronized with the evolving grasp.

Method

Graph-conditioned generation with geometry refreshed in the loop.

Object tokens, grasp tokens, and end-effector tokens meet inside a transformer generator conditioned by morphology-aware structure.

EAGG method overview diagram
EAGG encodes object geometry and end-effector geometry, injects topology-aware conditioning, and updates hand/gripper features throughout denoising.

Results

Cross-object, cross-hand, and hardware behavior.

Results use high-resolution paper figures where possible, with lightweight previews for generated grasp galleries.

Real-world EAGG hardware evaluation figure

Cross-regime grasp synthesis

Generated grasps span training, finetuned, and zero-shot end effectors across object categories.

EAGG ablation and quantitative analysis figure

Real-world evaluation

Hardware trials cover FreedomHand, Dahuan AG95, and SOARM 101 setups in physical scenes.

Allegro generated grasps across objects

Cross-object gallery

The same embodiment is applied to multiple clean objects with ranked generated grasps.

Generated mug grasps across end effectors

Cross-end-effector gallery

A fixed object is paired with different released hands and grippers to show embodiment transfer.

Resources

Paper, code, checkpoints, and dataset.

The project links stay compact and easy to scan, with citation separated below for direct copying.

Paper arXiv:2606.18092 Code Inference, training, visualization tools, assets, and configs. Checkpoints Released final and per-gripper model weights. MGG Dataset MultiGripperGrasp dataset and project reference.

Citation

Cite EAGG.

Use the BibTeX entry below if this work is helpful for your research.

BibTeX

@misc{niu2026eagg,
  title = {EAGG: Embodiment-Aligned Grasp Generation via Geometry-Aware Graph Conditioning},
  author = {Niu, Wanhao and Ke, Qiyan and Sun, Yuan and Sun, Hao and Xu, Jie and Ma, Muyuan and Hu, Ruiqi and Sun, Fuchun},
  year = {2026},
  eprint = {2606.18092},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  doi = {10.48550/arXiv.2606.18092},
  url = {https://arxiv.org/abs/2606.18092}
}