Multi-Scale Balanced Grasp Pose Detection for Robotic Sorting of Diverse Consumer Goods

The application of grasp pose detection technology is critical for improving robotic sorting efficiency in diverse electronic consumer goods, yet existing RGB-D based algorithms primarily optimize overall accuracy while neglecting significant performance imbalances across objects of varying scales, particularly for small-scale items. To address this limitation, we propose introduces a multi-scale consumer goods 6-DoF grasp pose balanced detection algorithm to overcome the multi-scale detection challenges. First, we develop a deep residual PointNet++ encoder that integrates depth information with hierarchical residual connections, enhancing global semantic feature extraction and mitigating information loss of small-scale consumer goods. Second, we design a multi-scale grasp width grouping feature extraction module that simultaneously captures multi-scale local geometric features under varying receptive fields through parallel dilated convolution branches, effectively resolving feature representation disparities across scales. Third, we introduce a multi-scale weight balanced loss function that dynamically adjusts optimization weights based on grasp width of consumer goods, addressing inherent data imbalance during training. Evaluated on the GraspNet-1Billion benchmark, our framework achieves remarkable improvements over GraspNet baseline, with accuracy gains of 18.71%, 9.35%, and 7.04% on small-scale objects across three test sets. Real-world experiments on a Franka Emika 7-DoF robot demonstrate a 91.79% success rate in cluttered multi-object scenarios. The robot grasping project and code are available at: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://upc-ghy.github.io/Franka-Grasp</uri>, and <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/upc-ghy/GraspBalance</uri>.