Active Inference for Control and Torque Disturbance Estimation: A Quantitative Single-Degree-of-Freedom Experimental Validation
Year: 2026
Type: preprint
Abstract: Residual torque components not captured by nominal dynamics degrade tracking and complicate disturbance compensation in torque-controlled robots. Active inference (AIF) updates state estimates and control actions jointly through free-energy minimization, but how faithfully its internal quantities represent residual torque disturbances has not been quantified directly. We experimentally evaluated an AIF-based controller for simultaneous position tracking and disturbance estimation on a single-degree-of-freedom torque-controlled testbed, where an unknown payload induced repeatable model mismatch, and compared it with a baseline controller and a nonlinear disturbance observer (NDOB)-based controller under matched sinusoidal- and step-tracking conditions. Estimation fidelity was evaluated against a residual torque reference constructed from torque sensing and nominal-model subtraction. Under nominal conditions, the NDOB-based controller tracked the sinusoid most accurately. Under payload-disturbed conditions, the AIF-based controller showed the smallest position root mean squared error (RMSE) in sinusoidal tracking and the smallest steady-state and integral absolute errors in step tracking, and its disturbance estimate had lower RMSE than that of the NDOB-based controller, by approximately 25% and 28% under sinusoidal and step inputs, respectively. These results suggest that AIF can improve tracking robustness while yielding an internal quantity consistent with the residual torque reference under controlled model mismatch.
Source: Research Square
Authors: Hyungseok Ryu, Pilwon Hur
Institution: Gwangju Institute of Science and Technology
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
Award: G9001342437
OpenAlex ID: W7225983552
