Quadruped robot locomotion based on deep learning rules

Terrestrial locomotion of robots employs two main methods: the use of wheels or tracks, and the use of articulated joints. Articulated locomotion offers advantages over wheels, such as lower energy consumption, access to irregular and rough terrains, better maneuverability in homes by climbing and descending stairs, and improved planning. Additionally, articulated robots are lighter, can be produced and prototyped with 3D printers, and offer greater flexibility in various environments. However, they also present some disadvantages, including complex mechanical structures, increased control complexity, and expensive actuators. This work presents the design and implementation of a reinforcement learning model based on Proximal Policy Optimization (PPO) for the locomotion of a 12-degrees-of-freedom quadruped robot that ensures stable trajectory. For this purpose, a reinforcement learning model based on TensorFlow and Gym was implemented and tested in the Pybullet simulation environment. With the correct adjustment of the model's hyperparameters, maximum stability in the robot's walking trajectory is achieved. During walking, the robot attains a smooth response curve when measuring its center of gravity. The application of reinforcement learning in this context shows the potential for advanced control techniques to address the complexities of articulated robots. By optimizing control strategies and leveraging modern simulation tools, this study demonstrates improvements in the stability and performance of quadruped robots, contributing to the development of more efficient and versatile robotic systems.