Leveraging Lightweight Neural Networks and Catastrophic Forgetting Mitigation for a Two-Wheeled Self-Balancing Robot

As artificial intelligence (AI) technology advances, its applications expand into various fields, including robotic motion control. Self-balancing robots commonly employ PID or LQR control and typically require extensive parameter tuning for proper execution. An emerging alternative is using deep reinforcement learning algorithms, leveraging on Lightweight Neural Networks (NNs) and Catastrophic Forgetting Mitigation, in robot control. This paper details the process and outcomes of training a Q-network within a virtual environment via a deep Q-network (DQN) algorithm, preventing the network from getting overtrained and deploying it on a microcontroller to achieve a self-balancing robot. Additionally, it describes methods for facilitating interactions between the self-balancing robot and the real-world environment during training.