Control cinemático de un robot manipulador mediante Damped Least Squares con amortiguamiento adaptativo estimado por red neuronal

Abstract

This paper presents the implementation of a kinematic controller based on the Damped Least Squares (DLS) method for 3D Cartesian trajectory tracking using a Dobot-CR3 manipulator. The primary objective is to robustly solve differential inverse kinematics in ill-conditioned and singular configurations while enforcing joint velocity and position constraints. The key contribution lies in deploying a multilayer perceptron (MLP) to adaptively predict the optimal damping parameter λ in real time. Rather than relying on rigid manipulability heuristics, the network is trained over 60,000 joint configurations using a supervised teacher grid search that optimizes a multi-objective cost function penalizing tracking errors, joint velocity peaks, actuator saturation, proximity to joint limits, and inverse amplification gains. The proposed neural controller is evaluated against three benchmarks under identical conditions: quasi-pseudoinverse (λ = 0), calibrated constant damping (λ = 0.0125), and classic Yoshikawa-based damping. Simulation results demonstrate that the neural DLS reduces peak joint velocity prior to saturation by 70.76 % (from 4.972 rad/s down to 1.454 rad/s) and decreases peak effective inverse gain by 80.06 % (from 62.082 down to 12.381) compared to λ = 0, completely eliminating actuator saturation events (from 16 saturated steps to 0). While the constant damping baseline achieved a slightly lower mean tracking error (0.377 mm vs. 0.455 mm for the MLP), the neural approach delivers the best global trade-off between tracking accuracy, mechanical stress, and numerical singularity avoidance.