BiGraspFormer:
End-to-End Bimanual Grasp Transformer

Kangmin Kim1, Seunghyeok Back2, Geonhyup Lee1, Sangbeom Lee1, Sangjun Noh1, Kyoobin Lee1,
1Gwangju Institute of Science and Technology (GIST), 2Department of AI Machinery, Korea Institute of Machinery & Materials (KIMM)
Paper arXiv Code

BiGraspFormer enables generate stable bimanual grasps for dual arm robot manipulation.

Abstract

Bimanual grasping is essential for robots to handle large and complex objects. However, existing methods either focus solely on single-arm grasping or employ separate grasp generation and bimanual evaluation stages, leading to coordination problems including collision risks and unbalanced force distribution. To address these limitations, we propose BiGraspFormer, a unified end-to-end transformer framework that directly generates coordinated bimanual grasps from object point clouds. Our key idea is the Single-Guided Bimanual (SGB) strategy, which first generates diverse single grasp candidates using a transformer decoder, then leverages their learned features through specialized attention mechanisms to jointly predict bimanual poses and quality scores. This conditioning strategy reduces the complexity of the 12-DoF search space while ensuring coordinated bimanual manipulation. Comprehensive simulation experiments and real-world validation demonstrate that BiGraspFormer consistently outperforms existing methods while maintaining efficient inference speed (< 0.05s), confirming the effectiveness of our framework.

Video

Method Overview

Method Overview Diagram

Our goal is to predict bimanual grasps from an object point cloud. This task is challenging as it involves a 12-DoF action space, doubling the complexity of single-arm grasping. Each grasp needs to achieve force-closure stability while both arms coordinate to avoid collisions, maintain torque balance, and ensure overall stability. To tackle this challenge, we introduce the (SGB) grasp generation scheme, which decomposes the prediction of bimanual grasps into three structured stages. First, generate diverse single grasp candidates under basic stability constraints. Second, select feasible grasp pairs by discarding collisions and ensuring balanced force distribution. Finally, refine these pairs into stable bimanual grasps using learned features from both the object and single grasps. This formulation explicitly enforces both individual grasp quality and dual-arm coordination, decomposing the complex 12-DoF search space into a sequence of more tractable subproblems.

Simulation Benchmark

Simulation Environments

Simulation Environment Diagram

Success rate and diversity at top K% in easy simulation environments

Simulation Easy

Success rate and diversity at top K% in hard simulation environments

Simulation Hard

As demonstrated in the tables, our method achieves the state-of-the-art performance in both easy and hard simulation settings. Notably, it also features the fastest inference speed among all compared methods, while showing a more significant improvement in success rate in hard settings compared to easy settings.

Real Robot Experiments

Real Robot Setting

Real Robot Setting

Dual-arm robotic system with two UR5e arms and two Azure Kinect RGB-D cameras, along with test objects and visualization of feasible bimanual grasps overlaid on object point clouds. Collision-free trajectory execution of selected grasp poses for grasping and lifting diverse objects including a stool, chair, and large box.

Real Evaluation

bluechair

toybox

White Shelf

Yellow Stair

Wood Stool

White Frame

Green Bin

Yellow Chair

BibTeX

@article{kim2025bigraspformer,
  title={BiGraspFormer: End-to-End Bimanual Grasp Transformer},
  author={Kim, Kangmin and Back, Seunghyeok and Lee, Geonhyup and Lee, Sangbeom and Noh, Sangjun and Lee, Kyoobin},
  journal={arXiv preprint arXiv:2509.19142},
  year={2025}
}