Human-in-the-Loop Reinforcement Learning for a Low-Cost Soft Myoelectric Prosthetic Hand

Approximately 65 million people live with some form of amputation, of which 38.7% involve the upper limb, with higher incidence in developing countries. This situation highlights the need for affordable and easily manufacturable hand prostheses that can improve users’ quality of life. Soft prosthetic hands are a promising alternative because they offer fewer moving parts and greater adaptability to anatomical variation; however, they also pose significant challenges, particularly the need for precise and efficient control. This work presents a Reinforcement Learning (RL) approach to control a soft myoelectric hand prosthesis from surface EMG (sEMG) signals within an embodied, human-in-the-loop formulation. A low-cost 4-DoF soft hand was 3D-printed from an open-source design using flexible materials. The RL controller was trained in two stages. Stage (i) was simulation pre-training, using a simplified dynamical model of the prosthesis and randomly sampled segments from an offline multi-participant sEMG–flexion dataset for concept validation and hyperparameter selection. Stage (ii) was fine-tuning on the physical hardware while preserving sensing–action timing to facilitate simulation-to-hardware transfer. Three reward formulations were evaluated, with the best results obtained using a hybrid combination of distance-to-target and discrete action-correctness terms. The controller was evaluated with three participants under controlled flexion/extension trajectories. Under these conditions, participant-level success rates were 86%, 74%, and 82% for participants 1, 2, and 3, respectively (overall approximately 81%). Success was defined as tracking the target trajectory below a root-mean-square error (RMSE) threshold of 440, calibrated with participant 1 and then kept fixed for the remaining participants. These results support the feasibility of RL-based control for a low-cost soft myoelectric hand under controlled flexion/extension conditions.