A Two-Stage Synergy-Based Control Framework for sEMG-Driven Robotic Grasping with Residual Reinforcement Learning

This thesis investigates surface electromyography (sEMG)-based control of a dexterous robotic hand and proposes a two-stage framework that combines supervised learning with residual reinforcement learning. In the first stage, eight-channel sEMG signals are decoded into six-dimensional synergy control signals. A comparison between a multilayer perceptron and a temporal convolutional network shows that the temporal model provides more accurate and consistent signal prediction and is therefore selected as the baseline decoder. In the second stage, a residual TD3 policy uses hand-state, contact, force, and holding feedback to apply bounded corrections to the baseline control signal. A zero-residual regularization term is introduced to limit unnecessary deviation from the decoded user intention and to improve training stability. The proposed method is first evaluated in a six-dimensional synergy-control task in MuJoCo, where it improves stable holding performance and reduces excessive contact forces, unsafe states, and control cost while preserving the main temporal structure of the baseline signal. The same residual-control principle is then adapted to a one-dimensional PowerGrasp interface using real-time sEMG recordings and human-operated hand trajectories reconstructed in MuJoCo. The results show similar improvements in holding stability, contact safety, force efficiency, and friction demand. Overall, the proposed framework separates intention decoding from task-level optimization and provides a practical approach for improving dexterous-hand grasping without replacing the user-generated baseline control signal.