Learning an Efficient Gait Cycle of a Biped Robot Based on Reinforcement Learning and Artificial Neural Networks

Programming robots for performing different activities requires calculating sequences of values of their joints by taking into account many factors, such as stability and efficiency, at the same time. Particularly for walking, state of the art techniques to approximate these sequences are based on reinforcement learning (RL). In this work we propose a multi-level system, where the same RL method is used first to learn the configuration of robot joints (poses) that allow it to stand with stability, and then in the second level, we find the sequence of poses that let it reach the furthest distance in the shortest time, while avoiding falling down and keeping a straight path. In order to evaluate this, we focus on measuring the time it takes for the robot to travel a certain distance. To our knowledge, this is the first work focusing both on speed and precision of the trajectory at the same time. We implement our model in a simulated environment using q-learning. We compare with the built-in walking modes of an NAO robot by improving normal-speed and enhancing robustness in fast-speed. The proposed model can be extended to other tasks and is independent of a particular robot model.