Accelerometer-in-the-loop safe learning control of mesh-order vibrations in cycloidal drives.
Cycloidal (RV-type) reducers are widely used in industrial robot joints due to their high torque density and low backlash, yet their multi-mesh transmission path produces structured, operating-point-dependent vibration components at the disc-mesh order and associated harmonics and sidebands. This paper presents a real-time, accelerometer-in-the-loop vibration suppression framework that reduces these components online while maintaining tracking performance within the bounds observed in our experiments and operating within predefined safety limits. A tri-axial accelerometer mounted on the reducer housing provides high-bandwidth vibration measurements from which order-synchronous, band-limited metrics are computed in streaming form. These metrics define both the optimization objective and vibration exposure constraints. The control architecture retains the vendor servo loops and adds a vibration-targeted layer combining a low-dimensional anti-resonance parameterization (adaptive notch shaping and narrowband feedforward cancellation aligned with the estimated mesh-order family) with a safety-certified contextual Bayesian optimization module that adapts the parameters as a function of operating context (speed, load proxy, and temperature proxy). A barrier-function-based safety filter runs at the servo rate to enforce constraint handling during operation; its effect is evaluated empirically through logged interventions and constraint statistics. Experimental evaluation on a cycloidal joint testbed across multiple speeds and load levels shows attenuation of the dominant mesh-order vibration component and its harmonics. Tracking accuracy and safety-related signals remained within preset limits during the tested operating conditions. The proposed approach provides a deployable pathway for online vibration minimization in cycloidal robot joints without requiring high-fidelity internal contact models, and its logged parameter trajectories and order-tracked metrics also offer a foundation for condition-aware adaptation over long-term operation.