A Deep Reinforcement Learning-based Application Framework for Conveyor Belt-based Pick-and-Place Systems using 6-axis Manipulators under Uncertainty and Real-time Constraints

Automated pick-and-place systems are seen as some of the most sensible commercial applications. However, developing such a system is still challenging because it requires not only high quality connected components but also that these connected components be synchronized to meet uncertainty and real-time constraints. This paper introduces a simulated conveyor belt-based pick-and-place system and proposes a joint learning algorithm to control manipulators for performing a sequence of pick and place tasks. The framework uses a deep reinforcement learning algorithm to teach an agent to pick objects that uncertainly and partial-observably appear on the conveyor. Moreover, this framework uses an Robot operating system (ROS) service to generate trajectories that navigate manipulators between objects and placements. Using an industrial-grade platform like ROS allows the generation of reliable path motions that meet various hardware and real-time constraints. We evaluate the effectiveness of our framework over a wide range of simulations against rule-based methods.