Model-free Bin-Picking: Food Processing and Parcel Processing Use Cases

Vision guided robots are enjoying growing success in industry and logistics, thanks to their adaptability to unstructured contexts and applications. The bin-picking problem is a prominent example in this trend. In typical bin-picking applications, a robot is guided to pick known rigid objects randomly placed inside a container. Given the objects’ CAD models, it is possible to accurately estimate the object pose and to perform the grasp synthesis in a closed form. Unfortunately, domains such as the food industry and package delivery require to manipulate polymorphic and deformable objects; moreover, valid CAD models are often not available. A direct bin-picking technology transfer is here infeasible, due to the variability in shapes, size, and appearance of the elements to be manipulated. In this abstract, we present the SPIWI and the EACHPack research projects: both have the goal to bridge the gap between the current and desired capabilities of vision guided robots, by developing new bin-picking methodologies and systems well suited to deal with food processing and package delivery products. We address the challenges related to the variability of shapes of the objects by solving the object pose estimation and grasp synthesis problems in an unified way, inside a state-of-the-art instance segmentation data-driven model. The former will be modified to explicitly deal with polymorphic shapes. The latter will be exploited both to provide an initial estimate of the position of goods, and to infer in advance some features (e.g., object weight), useful for efficient object placing. The proposed systems will be easily adaptable to a wide range of applications, thus greatly improving their potential impact.