Perception of Entangled Tubes for Automated Bin Picking

Abstract

Bin picking is a challenging problem common to many industries, whose automation will lead to great economic benefits. This paper presents a method for estimating the pose of a set of randomly arranged bent tubes, highly subject to occlusions and entanglement. The approach involves using a depth sensor to obtain a point cloud of the bin. The algorithm begins by filtering the point cloud to remove noise and segmenting it using the surface normals. Tube sections are then modeled as cylinders that are fitted into each segment using RANSAC. Finally, the sections are combined into complete tubes by adopting a greedy heuristic based on the distance between their endpoints. Experimental results with a dataset created with a Zivid sensor show that this method is able to provide estimates with high accuracy for bins with up to ten tubes. Therefore, this solution has the potential of being integrated into fully automated bin picking systems.

Detecting and Solving Tube Entanglement in Bin Picking Operations

“…The Entangled Tubes Bin Picking dataset from the authors' previous research [3] was expanded to include the files of all the point clouds acquired by the 3D scanner during the experiments. The dataset is available at https://github.com/GoncaloLeao/Entangled-Tubes-Bin-Picking-Dataset.…”

“…The four steps of this phase are: filtering, segmentation, cylinder fitting, and tube joining [3]. During the filtering step, all of the cloud points that do not belong to any of the tubes are removed.…”

“…The tubes on Set A are the same as those used on the conference article focused on the tube modeling phase [3]. They are straight plumbing pipes that were manually bent to have different shapes.…”

“…Since the focus of this research's experiments was accuracy, the simulation was set to run with a maximum step size of 0.001 and an update rate of 1000 Hz, which results in the simulation time running approximately as fast as real time. Moreover, unlike previous experiments conducted by the authors for the modeling phase [3], the whole, unfiltered point cloud was used to estimate the surface normals and curvatures, to ensure optimal accuracy for the modeling phase. Previous experiments had shown that downsampling the cloud used for normal estimation reduces the modeling time to around 1 s, which would reduce the average total processing time to around 5 s and 7 s for Sets A and B, respectively, when all picking attempts are considered.…”

“…This paper aims to present further advancements of the research regarding bin picking of entangled tube-shaped objects, by focusing on an industrial-inspired case study with radiator hoses from a vehicle's engine. Previously, a perception algorithm was developed to create a geometric model of the tubes from an input point cloud captured by a 3D scanner [3]. This article expands on the modeling algorithm by presenting a motion and grasp planning algorithm that allows a two-fingered gripper to pick, one-by-one, a set of freely curved tubes randomly arranged in a pile.…”