Title: Approximating Model Predictive Control via Behavior Cloning for 3-DOF Robotic Arms

Abstract: Although Model Predictive Control (MPC) is renowned for its robustness and stability guarantees, its high computational demands often hinder deployment in real-time applications. This study explores the use of Behavior Cloning to create surrogate policies that approximate MPC for the real-time operation of a three-degree-of-freedom (3-DOF) robotic manipulator. We establish a baseline control framework that integrates Inverse Kinematics with MPC and subsequently test various neural network structures—including traditional regression methods, Deep Multi-Layer Perceptrons (MLPs), and Recurrent Neural Networks (RNNs)—to identify computationally efficient alternatives. The research examines architectural trade-offs, stability implications, and the models' generalization potential. Through a combination of online and offline testing, we evaluate these surrogate policies based on their accuracy, speed, and alignment with the original MPC strategy. The findings indicate that Behavior Cloning significantly alleviates the computational load of MPC for 3-DOF manipulators, delivering an 84.98% success rate with relaxed tolerances while cutting inference latency by a factor of three. Intriguingly, static network architectures proved superior to temporal ones, suggesting that instantaneous state data is sufficient for this specific application. Nevertheless, a discrepancy in precision emerged under strict tolerance conditions, implying that while Behavior Cloning successfully captures the global optimal trajectory, additional investigation is required to reduce terminal steady-state errors.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC