Overview

My M.Sc. thesis, titled "Formation Control of Multi-Robot Systems Based on Learning Model Predictive Control", focused on advancing the control mechanisms for multi-robot systems operating in dynamic and uncertain environments. The research explored innovative methods to achieve coordinated behavior among multiple robots, emphasizing adaptability, computational efficiency, and robustness in the face of noise, disturbances, and unknown dynamics.

This work culminated in two significant publications that reflect the evolution of the proposed methodologies:

Publications

Title	Journal/Conference (peer-reviewed)
Control of Robots Using Convex QP LMPC and Learning-Based Explicit-MPC	IEEE Transactions on Industrial Informatics
Control of Multiple Robots Using LSTM-Based Model Predictive Control	International Conference on Robotics and Mechatronics

Abstracts

Publication	Abstract
Control of Robots Using Convex QP LMPC and Learning-Based Explicit-MPC	This article presents a novel method for quadrotor trajectory control utilizing cascade control and model predictive control (MPC). The proposed approach divides the control problem into a linear position controller, employing linear MPC with convex quadratic programming, and a nonlinear attitude controller, utilizing deep neural network-based MPC. Addressing the computational load challenges associated with online control, the hardware-in-the-loop (HIL) controller is tested to demonstrate its effectiveness in ensuring fast processing and suitability for online control. The stability of the proposed control strategies is analyzed, and simulation results using the HIL system validate the accurate tracking of desired trajectories. The findings highlight the functionality, reliability, and potential of the proposed approach for real-time applications in quadrotor control.
Control of Multiple Robots Using LSTM-Based Model Predictive Control	This paper proposes a Long Short-Term Memory- (LSTM-) based Model Predictive Control (MPC) design for the formation of multi-agent robots. The quadrotor model is considered nonlinear, and no simplifying assumptions are made, ensuring more reliable results and broader applicability of the proposed method. A cascade controller with a position controller as the main loop and an attitude controller as the inner loop is presented to address the formation control problem. An LSTM network is employed as the estimator to predict the future states of each quadrotor, while Differential Evolution (DE) optimizes input values to satisfy constraints and minimize tracking errors. Beyond its online learning capability, the proposed controller effectively handles environmental or structural disturbances, uncertainties, and noise. Additionally, a closed-loop stability analysis is performed. Finally, simulation examples are provided to demonstrate the effectiveness of the proposed control structure.

Key Differences

Aspect	LSTM-Based Model-Free MPC	Convex QP LMPC & Explicit-MPC
Core Design	Designed to handle noise, disturbances, and uncertainties without relying on known robot dynamics.	Developed as a low-computation controller focused on trajectory optimization and efficiency.
Computational Costs	High computational costs due to the use of LSTM networks for dynamic prediction.	Optimized for low computational demand, making it suitable for real-time applications.
Limitations	Struggles with maintaining low computational requirements for large-scale or real-time systems.	Challenges in handling disturbances, multi-rate systems, and other complex scenarios.
Applications	Ideal for environments with high uncertainties where model-free adaptability is critical.	Better suited for structured tasks where low computation is a priority over complex adaptability.

Conclusion

Both approaches represent important advancements in MPC for robotics, addressing distinct challenges in control. The LSTM-Based Model-Free MPC excels in handling uncertainties but demands significant computational resources, making it less practical for real-time applications. Conversely, the Convex QP LMPC and Explicit-MPC approach prioritizes computational efficiency, enabling real-time control but encountering limitations in managing disturbances and multi-rate systems. Together, these works lay the foundation for future research aimed at combining their strengths to achieve robust, adaptable, and efficient control strategies.