Automatic identification of mechanical parts for robotic disassembly using the PointNet deep neural network

Identification is the first step towards the manipulation of mechanical parts for robotic disassembly and remanufacturing. This paper presents a case study on the identification of objects from 3D scenes (point clouds) of mechanical components of automotive devices. The identification task is carried out through PointNet, a recently developed deep neural network system. PointNet is capable of identifying objects irrespective of their position and orientation in the point cloud. In this work, PointNet was used to recognise twelve instances of parts of different turbocharger models for automotive engines. The training instances consisted of different types of mechanical parts, as well as different models of the same type of part. Point clouds of partial views of the parts were created from CAD models using a purpose-developed depth-camera simulator. Different levels of sensor imprecision/noise were simulated. The results of the tests indicated that PointNet can be trained to recognise with good accuracy the various mechanical objects, and that its learning procedure is consistent and effective. In presence of sensor imprecision, the recognition accuracy in the recall phase can be increased adding some stochastic error to the training examples. The possibility of training twelve independent classifiers to be employed separately or in one ensemble classifier was also investigated. The accuracy results were comparable to those obtained using one classifier for all the parts.