Robotic peg-in-hole assembly is an essential task in robotic automation research. Reinforcement learning (RL) combined with deep neural networks (DNNs) lead to extraordinary achievements in this area. However, current RL-based approaches could hardly perform well under the unique environmental and mission requirements of fusion applications. Therefore, we have proposed a new designed RL-based method. Furthermore, unlike other approaches, we focus on innovations in the structure of DNNs instead of the RL model. Data from the RGB camera and force/torque (F/T) sensor as the input are fed into a multi-input branch network, and the best action in the current state is output by the network. All training and experiments are carried out in a realistic environment, and from the experiment result, this multi-sensor fusion approach has been shown to work well in rigid peg-in-hole assembly tasks with 0.1mm precision in uncertain and unstable environments.
View on arXiv