• artificial vision
• deep learning
• depth data
• industry 4.0
• motion control
• object detection
• RGB-D sensors
• robotic manipulation
• truncated signed distance fields

This work focus on the job of palletising and depalletising using robotic manipulators in the context of Industry 4.0 and transition to automation. More precisely, the process of unpacking a pallet requires an ample variety of tasks, which are the detection of the good, its localization in the space, and finally the picking of it, so that it is ready to be placed in the desired pose on a second pallet. The objective of this thesis is to find a way to determine the said pose in an automatic manner percepting the object in the space using the information generated by 3D visual sensors, in particular from RGB-D cameras, determining also how they are located in the space, i.e. their pose, and finally utilize this result in the subsequent motion control.
The work is subdivided in many steps, so the first concerns the implementation of a state-of-the-art convolutional neural network for object recognition. This one is exploited for the classification of a particular object belonging to a previously created database, giving in input 3D grids generated starting from depth data so that the type of the recognized object can be obtained in output. Then an extension of the network is required to detect many istances of the object, and to find how these are placed on more complex scenes and then set the strategy to pick them up; more in details, the algorithm performs a registration and then computes the position and the orientation of the objects, to provided them to the inverse kinematics part.
The validity and effectiveness of the method is shown both through simulation results and hardware experiments carried out using cameras to test the neural network and the whole algorithm with the robot.