A robotic wheel locally transforming its diameters and the reinforcement learning for robust locomotion

The implementation of the neural network has been paid attention in the autonomous operation of robots. In particular, it is efficient for a robot itself to learn the locomoting method to get over different obstacles on rough terrains. We are developing a robotic wheel that can locomote stably even on rough terrain, and introduce the reinforcement learning for the ability to robustly get over an obstacle. Our robot is able to locomote by utilising the extension and returning of the diameters by moving its centre of gravity. We study its mobility through four experiments, which are the testing of the locomotion on flat ground, the climbing over a step, controlling the robotic wheel by IMU, and the braking performance. After the learning, we verify the performance of getting over a step of 10 cm and 20 cm, which are equivalent to 25% and 50% of the wheel diameter, respectively.