Journal: Journal of Field Robotics

Article Title: An online hyper‐volume action bounding approach for accelerating the process of deep reinforcement learning from multiple controllers

doi: 10.1002/rob.22355

Figure Lengend Snippet: FIGURE 1 The outline of the proposed online action‐optimizer combined with the reinforcement learning (RL) method for a visual servoing (VS) application. The Twin Delayed Deep Deterministic Policy Gradient (TD3) agent was trained by using the Domain Randomization (DR) method. The Action Optimizer for improving Reinforcement Learning from multi‐Demonstrations (AORLD) method employs a combination of three different methods (position‐based visual servoing, image‐based visual servoing, and hybrid decoupled visual servoing) as demonstrators to accelerate training and enhance VS performance. Four different approaches have been used to constrain the action space of the agent and their results are compared together.

Article Snippet: The TD3 algorithm was implemented as a ROS node, and Matlab Reinforcement Learning Toolbox (MATLAB, 2021) was used to train policies.

Techniques:

Journal: Journal of Field Robotics

Article Title: An online hyper‐volume action bounding approach for accelerating the process of deep reinforcement learning from multiple controllers

doi: 10.1002/rob.22355

Figure Lengend Snippet: FIGURE 2 Structure of Action Optimizer for improving Reinforcement Learning from multi‐Demonstrations (AORLD) integrated with TD3 RL. The proposed block diagram in this study takes in current and desired features extracted from the vision sensor as inputs. Then, during each episode, the knowledge from hybrid decoupled visual servoing, position‐based visual servoing, and image‐based visual servoing approaches is utilized to restrict the action space. The joint velocity actions are then applied to the training environment, and the average rewards are computed accordingly.

Article Snippet: The TD3 algorithm was implemented as a ROS node, and Matlab Reinforcement Learning Toolbox (MATLAB, 2021) was used to train policies.

Techniques: Blocking Assay