This article presents a comprehensive study of a control model for legged robotic platforms, particularly hexapods, based on the application of deep reinforcement learning techniques. The relevance of employing artificial neural networks to form adaptive robot behavior in undefined conditions is substantiated, enabling greater flexibility and robustness in dynamic environments. The study includes a detailed analysis of modern simulation tools, including MuJoCo, PyBullet, Webots, Unity ML- Agents, Gazebo, and Nvidia Isaac Gym. Based on criteria such as computational efficiency, compatibility with popular deep learning frameworks, and scalability, the Nvidia Isaac Gym simulator is justified as the primary environment for agent training and simulation. A model was developed for the controlsystem that integrates data from multiple sensors – proprioceptive sensors, inertial measurement units (IMU), and stereo cameras. A data preprocessing pipeline is proposed, involving filtering, normalization, and the generation of a height map of the surrounding environment in the form of a fixed-size tensor. This approach enhances the generalization capabilities and stability of the trained neural network when transitioning from simulated to real-world environments. A supervisory module has been introduced to perform short-term prediction of neural network actions and validate their outputs against the platform’s physical constraints using a mathematical model of the robot. This real-time mechanism allows for the early detection of potentially hazardous behaviors and facilitates proactive mitigation of dangerous situations. The scientific novelty of the research lies in the comprehensive development of a hybrid control system that combines deep learning techniques with elements of classical control, thereby improving the reliability and practical applicability of the system in real-world scenarios. The practical significance of the work is determined by the scalability and adaptability of the proposed architecture for deployment in autonomous mobile robotic systems operating in complex, dynamic, and unpredictable environments. The conclusions highlight the importance of the hybrid approach that combines the strengths of classical control algorithms with the adaptability of deep learning. Future research directions include enhancing safety mechanisms and conducting real-world testing of the proposed system.
Matsiuk et al. (Wed,) studied this question.