Robotic Action-state Evaluation via Siamese Neural Network

Robotic imitation learning methods assist robots to operate in evolving and unconstrained environments. However, current robotic state representation imitation learning methods still must involve human experts to provide sparse rewards that indicate whether robots successfully complete tasks. However, enabling robots to make the action-state evaluation autonomously still remains a challenge, especially for multi-stage complex tasks. Therefore, in this work, we propose a novel Siamese neural network-based robotic action state evaluation system in an imitation learning system, so as to replace human experts in a multi-stage imitation learning process and improve the learning efficiency. One target learning footage is divided into several stages; for each stage, two Siamese network frameworks are created to assess the robotic action-states in terms of both movement and environment changes.