REINFORCEMENT LEARNING-BASED CONTROL SYSTEM FOR A TWO-WHEELED SELF-BALANCING MOBILE MANIPULATOR

Abstract

Reinforcement Learning (RL) provides a promising approach for autonomous control without explicit programming. However, training in real‑world environments is both time‑consuming and costly, while simulations, despite offering unlimited trials, can suffer from a sim‑to‑real gap. In this thesis, we develop an RL framework to train a custom Two‑Wheeled Self‑Balancing Mobile Manipulator to autonomously balance on its wheels and navigate toward randomly generated target points. The robot is designed using Fusion 360, Onshape, and NVIDIA Isaac Sim to accurately reflect real‑world parameters such as mass, dimensions, torque, damping, and velocity. We employ both velocity‑based and torque‑based control paradigms, using a curriculum learning strategy in Isaac Lab that progressively breaks down the overall control task into four subtasks: maintaining upright balance, orienting toward the target, moving toward it, and performing a soft landing. This structured approach minimizes training time and dataset size while enhancing learning efficiency. To validate the proposed methodology, we performed sim‑to‑sim and sim‑to‑real transfer experiments and compared our RL‑trained policies with a PID‑based controller. The results demonstrate that both control paradigms robustly achieve stabilization and target‐directed navigation, confirming that curriculum learning effectively accelerates training and bridges the sim‑to‑real gap for complex robotic tasks.