PS6D: Point Cloud Based Symmetry-Aware 6D Object Pose Estimation in Robot Bin-Picking

TL;DR

PS6D introduces a point-cloud-only, symmetry-aware framework for 6D object pose estimation tailored to industrial bin-picking. It combines an attention-guided feature extractor with a center distance sensitive translation loss and a symmetry-aware rotation loss, followed by a two-stage clustering pipeline to achieve accurate instance segmentation and pose voting for slender and multi-symmetric parts. The method demonstrates superior performance over state-of-the-art approaches on industrial datasets and achieves a 91.7% real-world bin-picking success rate, validating its practical applicability. The work also provides a dedicated PS6D dataset and deployment in Mech-Mind software, highlighting its potential for robust, textureless-object manipulation in manufacturing settings.

Abstract

6D object pose estimation holds essential roles in various fields, particularly in the grasping of industrial workpieces. Given challenges like rust, high reflectivity, and absent textures, this paper introduces a point cloud based pose estimation framework (PS6D). PS6D centers on slender and multi-symmetric objects. It extracts multi-scale features through an attention-guided feature extraction module, designs a symmetry-aware rotation loss and a center distance sensitive translation loss to regress the pose of each point to the centroid of the instance, and then uses a two-stage clustering method to complete instance segmentation and pose estimation. Objects from the Siléane and IPA datasets and typical workpieces from industrial practice are used to generate data and evaluate the algorithm. In comparison to the state-of-the-art approach, PS6D demonstrates an 11.5\% improvement in F$_{1_{inst}}$ and a 14.8\% improvement in Recall. The main part of PS6D has been deployed to the software of Mech-Mind, and achieves a 91.7\% success rate in bin-picking experiments, marking its application in industrial pose estimation tasks.

Figures (6)

• Figure 1: Existing algorithms encounter problems. (a) and (b) show the occurrence of missed detections when slender objects intersect, and the inability to accurately estimate pose; (c) shows the problem of large rotational regression bias for multi-symmetric objects. In (a), the red point cloud represents the real poses, while the green points represent the predicted centroids. In (b) and (c), the red point cloud represents the predicted poses, and the green point cloud represents the real poses.
• Figure 2: Architecture of PS6D. The network takes only the per-point position information as input. It performs feature extraction on the normalized point cloud, predicts translation and rotation, and finally undergoes two clustering stages and pose voting. This process ultimately achieves instance segmentation and pose estimation.
• Figure 3: Three models with different symmetries. (a) with d$_x$=d$_y$=0, d$_z$=180, and S=[ [ [1, 0, 0], [0, 1, 0], [0, 0, 1] ], [ [-1, 0, 0], [0, -1, 0], [0, 0, 1] ]], (b) without symmetry, (c) with infinite symmetry around the z-axis and no symmetry around the other axes.
• Figure 4: PS6D Dataset Object Models. Among them, (a), (c), and (d) are objects without symmetry, while (b) and (e) are objects with finite symmetry.
• Figure 5: Simulation visualization results of instance segmentation and pose estimation by PS6D on various objects. In the images above, different colors represent different instances, while in the images below, red point clouds represent predicted poses, and green point clouds represent ground truth poses.