An imitation learning approach for the control of a low-cost low-accuracy robotic arm for unstructured environments

We have developed an imitation learning approach for the image-based control of a low-cost low-accuracy robot arm. The image-based control of manipulation arms is still an unsolved problem, at least under challenging conditions such as those here addressed. Many attempts for solutions in the literature are based on machine learning, generally relying on deep neural network architectures. In typical imitation approaches, the deep network learns from a human expert. In our case the network is trained on state/action pairs obtained through a Belief Space Planning algorithm, a stochastic method that requires only a rough tuning, particularly suited to unstructured and dynamic environments. Our approach allows to obtain a lightweight manipulation system that demonstrated its efficiency, robustness and good performance in real-world tests, and that is reproducible in experiments and results, despite its inaccuracy and non-repeatable kinematics. The proposed system performs well on a simple reaching task, requiring limited training on our quite challenging platform. The main contribution of the proposed work lies in the definition and real-world testing of an efficient controller, based on the integration of Belief Space Planning with the imitation learning paradigm, that enables even inaccurate, very low-cost robotic manipulators to be actually controlled and employed in the field.