Abstract
Legged robots, especially quadruped robots, have proven to be very useful for different applications due to their ability to locomote over different terrains. However, this type of robot requires significantly more energy compared to other types of simpler robots. In this work, we implement an energy-efficient locomotion controller based on Deep Reinforcement Learning to follow a desired velocity command minimizing power consumption, and for this we trained a Proximal Policy Optimization (PPO) agent in Isaac Lab simulator, and all the modeling of the environment for training and controlling the robot is presented. We also implement a controller based on Central Pattern Generator so we can compare the power consumption of both controllers. We found that our trained policy learned an efficient gait for locomotion and can spend up to 55.8% less energy compared to the CPG controller in the best case.