Data-Driven Compensation of Non-Geometric Joint Errors: A Comparative Study in Configuration and Task Spaces

초록

This paper presents two data-driven compensation strategies for improving pre-contact end-effector accuracy in serial leg mechanisms subject to non-geometric joint errors. Mechanical clearance and assembly tolerances introduce parasitic rotational deviations that are not captured by motor encoder measurements, leading to significant foot-position errors before ground contact. To address this issue, we investigate (1) a configuration-space compensation approach that estimates internal play angles and incorporates them into the forward kinematics, and (2) a task-space compensation approach that directly learns the end-effector residual error. Ground-truth link states are obtained using a motion-capture-based calibration pipeline, and both strategies are trained using motor-side measurements including joint angles, velocities, and currents. Experimental evaluation under varying configurations shows that configuration-space compensation is more effective in mechanically stable conditions, while task-space compensation demonstrates greater robustness under larger structural uncertainties. The results highlight the trade-off between physically interpretable correction and end-to-end residual learning, and provide insight into selecting appropriate compensation strategies for legged robotic systems.

키워드