Mecanum-Wheeled Robot Control Based on Deep Reinforcement Learning

The use of mecanum wheels for mobile robots has led to more flexibility in terms of the robots' navigation ability. This usage, however, leads to a more complex control system which mainly relies on a complete knowledge of the robot's kinematic model. This research aims to show that reinforcement learning can help develop control systems for robots with mecanum wheels. Since reinforcement learning does not require the robot's model explicitly to generate the control action, it can deal with the uncertainty and nonlinearity of the robot model by learning the action directly from interaction with the environment. In the paper, we proposed a novel reward function specifically designed for mecanum-wheeled robot control. The reward function is used to train the reinforcement learning algorithm to produce control signal for mecanum-wheeled robot. The results demonstrate the promising capability of reinforcement learning in controlling mecanum-wheeled robot towards a target location while maintaining its orientation.