多尺度物体杂乱堆叠场景下的机器人抓取位姿检测

Robotic Grasp Pose Detection in Cluttered Scenes with Multi-scale Object Stacking

摘要:

针对多尺度物体杂乱堆叠场景因物体边界模糊、相互遮挡,易导致机器人稳定抓取位姿检测困难的问题,提出一种多尺度特征融合的机器人稳定抓取位姿检测方法。 首先,设计一种结合 KD-tree 加速欧氏聚类与高斯采样的两阶段点云预处理方法,提取稳定抓取目标点并抑制环境噪声。 其次,通过多尺度圆柱采样与位置编码联合提取抓取中心的局部几何特征,并引入多头自注意力机制提取全局上下文特征,提升抓取稳定性与抓取检测精度。 最后,在 GraspNet-1Billion 数据集上进行抓取位姿准确度评估实验,并在 UR10 机器人上进行真实场景下的抓取实验;实验结果表明,所提方法在 3 类别数据集中的平均精度(AP)分别为 73. 46% 、65. 64% 和 26. 29% ,随机杂乱场景下机器人的平均抓取成功率为 85. 71% ,三层堆叠实验场景下机器人的抓取成功率为 81. 40% ,验证了所提方法的有效性。

Abstract:

Robotic grasp pose detection in cluttered scenes with multi-scale objects remains challenging due to severe occlusion and ambiguous object boundaries, which often leads to unstable grasp predictions. To address this problem, a robotic stable grasp pose detection method was proposed based on multi-scale feature fusion. Firstly, a two-stage point cloud preprocessing strategy combining KD-tree- accelerated Euclidean clustering and Gaussian sampling was designed to extract stable grasp target points while effectively suppressing environmental noise. This preprocessing step improved the reliability of candidate grasp points in cluttered and stacked scenes. Secondly, local geometric features around grasp centers were extracted through multi-scale cylindrical sampling combined with positional encoding, enabling the representation of grasp-relevant geometry under different spatial scales. On this basis, a multi-head self-attention mechanism was introduced to capture global contextual features among grasp candidates, which further enhanced grasp stability and improved grasp pose detection accuracy under occlusion. Finally, extensive experiments were conducted on the GraspNet-1Billion dataset to evaluate grasp pose accuracy, along with real-world grasping experiments on a UR10 robotic platform. Experimental results showed that the proposed method achieved average precision (AP )values of 73. 46% , 65. 64% , and 26. 29% on the Seen, Similar, and Novel subsets, respectively. In real-world experiments, the robot attained an average grasp success rate of 85. 71% in random cluttered scenes and 81. 40% in three-layer stacked scenarios. These results demonstrated the effectiveness of the proposed method for stable grasp pose detection in cluttered scenes with multi-scale objects.