1.
Zvonarev I. S., Karavaev Y. L.
Deep reinforcement learning for mobile robot control for motion toward a given position
Robotics and Autonomous Systems, 2026, vol. 200, 105392, pp.
Abstract
This work addresses the implementation of multiple neural-network architectures and reward-assignment schemes within the Deep Deterministic Policy Gradient (DDPG) algorithm for controlling a differential-drive mobile robot in a target-reaching task. The study includes hyper-parameter optimization and evaluates algorithm performance in both simulation and real-world deployment. A bespoke simulation environment, developed in compliance with the principles of the OpenAI Gymnasium framework, was used for pre-training; the environment provides a faithful model of the differential-drive robot kinematics, ensuring realistic training conditions. A statistical comparison was carried out between our modified DDPG implementation and the baseline classical DDPG provided by the stable_baselines3 library. The optimized model was subsequently transferred to a physical robot prototype, with a motion-capture system ensuring precise positional feedback. Through a series of experiments, different network architectures were systematically evaluated, and the most effective reward-assignment policy per training episode was identified. The results demonstrate that architectural design choices and reward shaping have a substantial impact on performance in mobile-robot control, and highlight the necessity of careful hyperparameter tuning of the network.
| Citation: |
Zvonarev I. S., Karavaev Y. L., Deep reinforcement learning for mobile robot control for motion toward a given position, Robotics and Autonomous Systems, 2026, vol. 200, 105392, pp. |
| DOI: |
10.1016/j.robot.2026.105392 |
Journal Info
