A Unified Hybrid Control Architecture for Multi-DOF Robotic Manipulators

Multi-degree-of-freedom (DOF) robotic manipulators exhibit strongly nonlinear, high-dimensional, and coupled dynamics, posing significant challenges for controller design. To address these issues, this work proposes a unified hybrid control architecture that integrates model predictive control (MPC) with feedback regulation, together with a stability analysis of the proposed scheme. The proposed approach mitigates the optimization difficulty associated with high-dimensional nonlinear systems and enhances overall control performance. Furthermore, a hardware implementation scheme based on machine learning (ML) is proposed to achieve high computational efficiency while maintaining control accuracy. Finally, simulation and hardware experiments under external disturbances validate the proposed architecture, demonstrating its superior performance, hardware feasibility, and generalization capability for multi-DOF manipulation tasks.

💡 Research Summary

The paper addresses the longstanding challenge of controlling multi‑degree‑of‑freedom (DoF) robotic manipulators, whose dynamics are highly nonlinear, high‑dimensional, and strongly coupled. Conventional feedback controllers (PID, ADRC, H∞, adaptive, sliding‑mode) lack systematic optimization and scalability, while Model Predictive Control (MPC) offers a principled way to handle constraints and performance objectives but becomes computationally prohibitive for high‑DoF systems. To overcome these limitations, the authors propose a unified hybrid control architecture that tightly integrates a feedback regulator with an MPC layer, and they further replace the online MPC optimizer with an offline‑trained machine‑learning (ML) torque emulator.

The feedback component extracts a rich error vector E(t) containing position, velocity, acceleration, integral, filtered, and disturbance‑estimated terms. A generalized mapping ϕ(E) produces a set of candidate torques, weighted by a selection vector w_sel. These candidates are expanded over a prediction horizon using temporal weight matrices W_pre, forming a torque preview sequence that serves as the initial guess for the MPC optimizer. The MPC problem minimizes a weighted sum of tracking error and control effort subject to the robot’s dynamic model (computed via a Recursive Newton‑Euler Forward Dynamics Algorithm, RNEFDA) and state‑and‑input constraints. The authors derive sufficient Lyapunov‑based conditions guaranteeing local asymptotic stability of the closed‑loop system composed of feedback, MPC, and the ML emulator.

To achieve real‑time feasibility, the optimal torque law generated by the MPC is approximated offline by a deep neural network. An adaptive sampling strategy is employed to collect high‑quality state‑torque pairs across the robot’s workspace and under various external disturbances. The trained network, termed the torque emulator, predicts the optimal torque sequence at runtime with negligible latency, thereby eliminating the heavy online optimization burden while preserving the performance of the original MPC.

Extensive validation is performed on a 6‑DoF serial manipulator. In simulation, the hybrid controller outperforms baseline PID, ADRC, and a conventional nonlinear MPC in terms of root‑mean‑square tracking error (over 40 % reduction) and torque smoothness, even when subjected to sudden payload changes and impact disturbances. The ML‑based implementation achieves the same performance with a computational speed‑up of an order of magnitude, meeting a 2 ms control cycle. Hardware experiments corroborate these findings: the robot tracks complex trajectories under external forces with high accuracy, and the torque emulator runs in real time on an embedded processor without destabilizing transients.

The contributions are threefold: (1) a novel hybrid architecture that unifies feedback regulation and MPC for arbitrary‑DoF manipulators, together with rigorous stability analysis; (2) an ML‑based torque emulator with adaptive data collection that bridges the gap between model‑based optimal control and real‑time implementation; (3) comprehensive simulation and experimental evidence demonstrating superior accuracy, robustness, and computational efficiency. The work opens avenues for deploying high‑performance optimal control in collaborative robotics, flexible machining, and precision assembly, while future research will focus on extending global stability guarantees, online learning updates, and broader generalization to unstructured tasks.