A Closed-Loop Multi-perspective Visual Servoing Approach with Reinforcement Learning

The article discusses a novel learning-based multi-perspective visual servoing framework that uses reinforcement learning to estimate robot actions from latent space representations of visual states. The approach was tested in a Gazebo simulation environment with connection to OpenAI/Gym. It demonstrated a high success rate, outperforming the Direct Visual Servoing algorithm. The researchers highlight the limitations of traditional visual servoing methods and the benefits of their new approach, particularly in complex industrial scenarios.

Publication date: 29 Dec 2023
Project Page: 10.1109/ROBIO58561.2023.10354958
Paper: https://arxiv.org/pdf/2312.15809