Physics-Informed Learning-based model-free computed torque control for unknown nonlinear rigid dynamic systems

مقالات آماده انتشار

نوع مقاله : مقاله پژوهشی

نویسنده

آرمان مردانی

گروه طراحی جامدات- دانشکده مهندسی مکانیک- دانشگاه صنعتی نوشیروانی بابل

10.22067/jacsm.2026.95250.1354

چکیده

This paper presents a new model-free control strategy for nonlinear rigid-body systems with unknown or highly complex dynamics. Unlike traditional Computed Torque Control (CTC), which depends on precise analytical models, this approach uses real-time data to estimate system behavior, eliminating the need for prior modeling. The method operates in two key stages. First, it dynamically estimates the Jacobian matrix by applying small perturbations to control inputs, linking task-space motion to joint torques. This Jacobian maps a virtual spring-damper force from the workspace to joint space, ensuring stable trajectory following. Second, a constrained Recursive Least Squares (RLS) algorithm learns a simplified linear representation of the system’s nonlinear dynamics, accounting for Coriolis, centrifugal, and gravitational effects. Crucially, this learning process incorporates physical constraints, maintaining energy conservation and other essential rigid-body properties. data-driven mechanisms. The result is a real-time compensator for nonlinearities that preserves stability and performance. Tests on a robotic manipulator confirm improved tracking accuracy and robustness over conventional CTC, especially in cases with unmodeled dynamics or parameter uncertainties.

کلیدواژه‌ها

Numerical Jacobian Estimation

Perturbation-Based Learning

Nonlinear rigid-body systems

dynamic systems approximation

موضوعات

• رباتیک

---