Rapid speed control of traveling-wave ultrasonic motors using sensorless deep reinforcement learning

Traveling-wave ultrasonic motors (TWUSMs) offer high torque density and low rotor inertia. With a mechanical time constant of only 1–2 ms, these motors possess inherently fast response capability; however, conventional drivers typically achieve ∼50 ms startup. This study demonstrates speed control using deep reinforcement learning (DRL), achieving rise times of 0.8–1.6 ms. A Soft Actor-Critic agent was trained directly on hardware using driving current phasors as state feedback, enabling neural-network-based control at a ∼400 µs control period. The trained controller achieved steady-state RMSE of 4.82 rpm across target speeds of 50–250 rpm. Compared with a conventional driver, the proposed approach is 25 times faster and 2.2 times more accurate, while requiring approximately 35 s of motor operation for training. In addition, the rotary encoder is used only during training; inference relies solely on electrical measurements, enabling system miniaturization for compact or mechanically constrained applications.