6-DoF Grasp Detection in Clutter with Enhanced Receptive Field and Graspable Balance Sampling

Hanwen Wang, Ying Zhang, Yunlong Wang, Jian Li

Abstract

Robotic manipulation in cluttered environments remains a significant challenge, particularly when dealing with small-scale objects. This paper introduces a novel 6-DoF grasp detection network that addresses two critical bottlenecks: feature loss due to fixed receptive fields and sampling imbalance in 3D point clouds. By utilizing Multi-radii Cylinder Grouping (MrCG) to expand the receptive field and a 3D segmentation-based Graspable Balance Sampling (GBS) method, the proposed system achieves a 10% improvement in Average Precision (AP) on the GraspNet-1Billion benchmark and demonstrates high success rates in real-world validation.

Introduction

In robotic manipulation, grasp detection involves determining the optimal pose for a gripper to interact with an object. While deep learning models have become proficient at grasping large, isolated objects, they frequently fail in cluttered scenes containing small items. This failure is primarily attributed to two factors:

  1. Feature Loss: Standard neural networks often employ fixed-size receptive fields that either miss the fine details of small objects or lack the necessary context.
  2. Sampling Imbalance: Common techniques like Farthest Point Sampling (FPS) naturally favor large objects with high point densities, causing the network to effectively ignore smaller objects that represent a fraction of the total data.

Methodology

Enhanced Receptive Field (ERF)

To capture both local geometry and global context, the researchers developed the Multi-radii Cylinder Grouping (MrCG) module. Unlike traditional spherical grouping, MrCG uses four distinct radii to group points in a cylindrical volume, which better aligns with the geometry of a two-finger gripper. This allows the network to perceive fine-grained details, such as the handle of a pair of scissors, while maintaining awareness of the surrounding environment. To mitigate noise from the expanded receptive field, a Passive Attention (PA) module is integrated to prioritize relevant features.

Grasp detection on small-scale components like scissor handles and spoon grips using the enhanced receptive field.

Graspable Balance Sampling (GBS)

To solve the sampling imbalance, the team integrated a pre-trained 3D segmentation network. During inference, the system identifies individual object instances within the scene. The Graspable Balance Sampling method then ensures that an equal number of potential grasp points are sampled for every identified object, regardless of its physical size. This ensures that a tiny screw receives the same computational attention as a large bowl.

Experimental Results

Benchmark Performance

The model was evaluated on the GraspNet-1Billion dataset, which contains 190 scenes with dense annotations. The proposed method achieved state-of-the-art results, showing a 10% AP improvement over previous methods like GSNet. In the most challenging "Novel" object scenarios, the $AP_{0.4}$ metric increased from 14.23% to 20.23%.

Successful 6-DoF grasp candidates for small objects (Class 1 and Class 2) in cluttered scenes enabled by balanced sampling.

Simulation and Real-World Validation

In PyBullet simulations using objects from the YCB dataset, the model achieved a 95% success rate in scenes composed entirely of small-scale objects. Real-world testing with an AUBO-i5 robot arm further validated the approach, maintaining an 83% success rate in complex clutter.

Discussion and Conclusion

The study highlights that 6-DoF grasping is not a scale-invariant problem; networks must be "scale-aware" to handle clutter effectively. Using segmentation as a sampling guide proves to be a powerful strategy for overcoming data imbalance in 3D point clouds.

Despite these gains, the system still faces challenges with flat or thin objects where depth cameras fail to capture complete geometry. Future research will explore the integration of image-based instance segmentation and "semantic grasping" to further refine the robot's understanding of task-specific grasp points.

Summary updated

BibTeX
@inproceedings{wang2024dof,
  author = {Hanwen Wang and Ying Zhang and Yunlong Wang and Jian Li},
  title = {{6-DoF Grasp Detection in Clutter with Enhanced Receptive Field and Graspable Balance Sampling}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2024},
  doi = {10.1109/iros58592.2024.10802025},
}
Figure