Analýza pokročilých řídicích algoritmů pro chůzi čtyřnohých robotů se zaměřením na užití v simulaci

Abstract

The field of robotics has experienced a surge of advanced robots entering the market in recent years. Many robots today can perform tasks that were unimaginable only a few years ago. Among the various types of robots—such as drones, bipedal robots, or hexapods —four-legged robots have gained significant attention due to their ability to navigate complex terrains and perform dynamic movements.. In this work, we explore state-of-the-art methods for controlling quadrupedal robots. Specifically, we focus on implementing four walking controllers for the Anymal D robot, one of the most advanced four-legged machines available. These controllers include a model predictive controller (MPC) in combination with a whole-body tracking controller, a reinforcement-learning-based controller, and two hybrid controllers that combine an MPC controller with a neural-network based tracking controller. We verify the functionality and performance of these controllers in the Gazebo simulator, evaluating their ability to effectively traverse various types of terrains. Finally, we develop a context-aware strategy that dynamically switches between two of the implemented controllers based on the available information.The field of robotics has experienced a surge of advanced robots entering the market in recent years. Many robots today can perform tasks that were unimaginable only a few years ago. Among the various types of robots—such as drones, bipedal robots, or hexapods —four-legged robots have gained significant attention due to their ability to navigate complex terrains and perform dynamic movements. In this work, we explore state-of-the-art methods for controlling quadrupedal robots. Specifically, we focus on implementing four walking controllers for the Anymal D robot, one of the most advanced four-legged machines available. These controllers include a model predictive controller (MPC) in combination with a whole-body tracking controller, a reinforcement-learning-based controller, and two hybrid controllers that combine an MPC controller with a neural-network based tracking controller. We verify the functionality and performance of these controllers in the Gazebo simulator, evaluating their ability to effectively traverse various types of terrains. Finally, we develop a context-aware strategy that dynamically switches between two of the implemented controllers based on the available information.