Demonstration of learned behavior of an autonomous mobile robot as learned by a digital twin, in various stages of the learning process.
Keywords: Deep reinforcement learning, autonomous mobile robots, AGV, action selection, digital twin, simulation, robust behavior, RAAK Let’s Move IT project
One of the challenges of modern robotics is a correct robust behavior in unstructured/unpredictable environments. A recent trend is to deploy AI-techniques, such as deep reinforcement learning. Instead of manually designing the correct robot behavior, we choose to train the robot behavior by providing feedback, i.e., rewards. As learning can be done much more efficiently in a simulated world, the technique a so-called digital twin is applied. The virtual counterpart of the robot matches the real one, as closely as possible. This should enable effective transfer of the learned behavior from the simulated robot to the real one. Results of the approach will be shown for different scenarios in the workfloor with Automated Guided Vehicles (AGV).