Integrative Advances in Deep Reinforcement Learning for Electromechanical Systems Optimization and Autonomous Navigation

Project Manager: Zholtayev Darkhan Muratovich

Funding Source: GF of Projects for the youngest scientists «Zhas Galym»

Goal: The goal is to improve deep reinforcement learning (DRL) algorithms using fluid neural networks (FNN) and large language models (LLM) to create more adaptive and efficient DRL systems. These algorithms will be tested on complex electromechanical systems, the efficiency of the system and its adaptability in dynamic environments.

Years of implementation: 2024–2026

Funding amount: 29,858,850 tenge

1. Development of an autonomous navigation system based on TD3 and a depth camera;
2. Design and optimization of the reward function;
3. Improving the simulation environment for training and stress testing.

A system for autonomous navigation of a mobile robot was developed using the TD3 deep reinforcement learning algorithm and a depth camera that provides three-dimensional perception of the environment. An effective reward function was designed that takes into account the distance to the target, obstacle avoidance, and smoothness of movement, which allowed for faster training and increased stability of the agent’s behavior. The simulation environment was also improved by adding dynamic obstacles and realistic conditions necessary for stress testing and increasing the generalization ability of the model.

1. A conceptually validated and improved algorithm using deep reinforcement learning in electromechanical systems was developed using simulation.
2. A validated and improved deep reinforcement learning algorithm was obtained using simulation and its implementation on hardware was started.