Curriculum-organized Reinforcement Learning for Robotic Dual-Arm Assembly

Modern manufacturing heavily relies on robotic systems, yet collaborative assembly executed by two or more robot arms presents several challenges. Dealing with tight manufacturing tolerances demands a flexible and efficient methodology, empowering robots to handle tasks requiring precision and fine motor skills. In this work, as an example of this we formulate a peg-in-hole task involving two Franka Emika Panda robots. For these we employ a hierarchical control architecture. Our approach involves planning feedback-based trajectories for the robots using a reinforcement learning agent, which are transmitted to low-level impedance controllers on each robot. To facilitate the learning process, we structure a training procedure as a reverse curriculum, incorporating domain randomization. Upon completion of the training, we evaluate inferences from the controlled system within a simulated environment. Our analysis delves into the emerging process characteristics resulting from the curriculum parameters. Episode length and minimum reward threshold imply process time and its variance, which have been subject to great uncertainties in previous work.