Transparent Object Classification and Location Using MmWave Radar Technology for Robotic Picking

Abstract

Detection of transparent objects for classification or localization are among the most challenging tasks in flexible automation areas for highly personalized manufacturing. Additional issues arise when also applying the concepts of Lights-out-Manufacturing. The need to make processes more flexible and efficient are some of the key objectives of Industry 4.0, with automation going as far as removing all personnel from factory floors. In this sense, this paper presents a prototype tool proposing the use of mmWave (millimeter Wave) radars to detect, classify and locate objects, especially transparent ones, by evaluating two versions of radars produced by Texas Instruments. The best approaches to acquire data, classify and locate the objects with mmWave radars are explored using a combined solution of a robotic system and Deep Neural Networks (DNN) to process the cloud points. Using a total of 12 scanning routines, 6 showed more than 80% of detection accuracy, 2 near 40%, 1 above 70% and 3 were not able to be execute due to physical limitations. The classification of stationary objects showed limited results. Motion variations on object position shows an accuracy decreased, to around 40%. The velocity changes were also assessed, revealing that at slower velocities (0.03 m/s) the accuracy increases above 80%. The final system evaluation was executed in two approaches with raw data direct from sensor and normalized data around the axis coordinate, showing similar and promising results in both approaches. The localization did not show the best results, although improvement to the methodology is suggested.