Learning-Based-Controls-2025

MECHENG 6193: Spring 2025

Course Description:

This course explores the synergy between control theory and machine learning, enabling students to design learning-based control systems ensuring adaptability, robustness, stability, and safety. The course focuses on integrating data-driven methods with traditional control techniques to create innovative solutions for real-world applications such as autonomous systems, robotics, advanced manufacturing systems, chemical systems, etc. It applies theoretical concepts to practical engineering problems, incorporating research from past and ongoing projects into the course material, assignments, and an open-ended project. The methods aim to enhance adaptability and decision-making in dynamical systems including autonomous vehicles and robotics.

Instructor:

Dr. Qadeer Ahmed (ahmed.358@osu.edu)
Assistant Professor
Department of Mechanical and Aerospace Engineering
Department of Electrical and Computer Engineering
Department of Integrated Systems Engineering
Affiliated Faculty, Center for Automotive Research
Smart Vehicle Concepts Center
Sustainability Institute
Institute of Cybersecurity and Digital Trust

Course Assistant:

Dr. Sidra Ghayour Bhatti (bhatti.39@osu.edu)
Research Associate Engineer
Center for Automotive Research (CAR)
The Ohio State University (OSU)

Course Objectives:

Develop the motivation for bridging the gap between conventional control theory and learning techniques.
Review the fundamentals of control systems including feedback control, state-space modeling, LQR, and stability analysis.
Review the development and solution of optimal control problems using constrained and unconstrained optimization techniques.
Review the fundamentals of machine learning (ML) and develop understanding of advanced ML models, including physics-informed ML, while exploring their applications in dynamical systems and controls.
Develop understanding of multi-agent system (MAS), traditional reinforcement learning (RL) algorithms (model-based and model-free), deep RL algorithms, and their use in control-oriented applications.
Review Model Predictive Control (MPC) and its variants and develop understanding of Learning-based MPC (LMPC) for dynamical systems with a focus on integrating safe sets for secure learning.
Develop an understanding of Gaussian process regression (GPR) and its integration with LMPC, focusing on how GPR can be used to handle uncertainties within the LMPC framework.
Develop the learning-based controller and perform its stability analysis using quadratic constraints.
Develop an understanding of system identification for dynamical systems using Neural ODEs and Generative Adversarial Networks (GANs).
Enhance programming and mathematical analysis skills.

Syllabus:

The class syllabus can be found here Download syllabus (PDF)

Products:

Software: MATLAB 2025a, 2025b; Simulink; Control System Toolbox™; Reinforcement Learning Toolbox™; System Identification Toolbox™; Simscape™; Deep Learning Toolbox™; Python: PyTorch, TensorFlow/Keras, Stable Baselines3, RLlib, Gymnasium, PettingZoo ; CARLA; SUMO (traffic); etc.

Awesome Learning based Controls:

Learning-Based Controls is an emerging field at the intersection of control theory for dynamic systems and modern machine learning (deep learning and reinforcement learning), and this repository curates key resources, topics, tutorials, research articles, applications, and tools in that area.

Awesome Learning-Based Controls repository: GitHub link

Projects:

Freeform Surface Reconstruction and Adaptive Sampling through Machine Learning for Applications in Minimally-Invasive In-Vivo Bioprinting
Double DQN Energy Management strategy for a Series Hybrid Agricultural Tractor
Industrial Scale Reverse Osmosis Modeling and Control
Deep Koopman Operator Autoencoder Learning for Pendulum System
SafeNN: A Neural Network Controller for Safe Autonomous Driving
A Neural-Forecasting-Driven MPC Framework for Two-Intersection Traffic Signal Control
Physics Informed Neural Network based Parameter Estimation of Permanent Magnet Synchronous Machine
Energy-Efficient Powertrain Control Using Reinforcement Learning with Battery Aging Integration
Temporal Dependency Based Soft Actor-Critic for Engine Control in Series Hybrid Electric Vehicles
Bipedal Locomotion on Terrains with Restricted Footholds

MECHENG 7194: Autumn 2025 (Advanced Version)

Course Description:

This advanced course bridges control theory and modern learning (AI/ML/RL), covering model-free/model-based RL, offline vs online learning, transformer-based and MPC-integrated methods, and meta-RL with a strong emphasis on safety (ISO/PAS 8800). It introduces RL for cost-function learning, controller gain tuning, and safe MPC alongside advanced architectures such as liquid neural networks, graph neural networks/GATs, foundation and world models. It also covers GANs, physics-informed techniques, and Koopman-based system identification, together with contraction theory and region-of-attraction concepts for stability and safety. Explainability, interpretability, and trustworthiness of black-box ML/RL controllers are emphasized through theory and project-based applications in autonomous and cyber-physical systems.

Instructor:

Course Assistant:

Dr. Sidra Ghayour Bhatti (bhatti.39@osu.edu)
Research Associate Engineer
Center for Automotive Research (CAR)
The Ohio State University (OSU)

Course Objectives:

Develop the motivation for bridging the gap between conventional control theory and learning techniques.
Understand the concept of ISO/PAS 8800 to ensure the safety of learning-based control solutions.
Explore and implement advanced neural architectures including Spiking NN (SNN), Liquid NN (LNNs), graph NNs (GNNs), etc. for control tasks.
Develop the understanding of Foundation and World models including Artificial General intelligence (AGI) for dynamical system and controls.
Understand and apply advanced RL methods (e.g. DDQN, DTQN, TD3, DDPG, SAC, PPO, TADPO, GRPO), including model-free and model-based, online versus offline, on-policy versus off-policy RL, knowledge distillation, Transformer-based RL, MPC-integrated RL, and meta-RL.
Explore RL for learning cost functions and constraints in MPC, gain tuning for controllers (PID, LQR), and safe RL with MPC.
Utilize GANs, physics-informed methods, and Koopman operator for accurate system identification and dynamics modeling.
Apply control theories like contraction theory and region of attraction (ROA) to ensure stability, robustness, or safety in learning-based systems.
Develop the learning-based controller and perform its stability analysis using quadratic constraints and ROA.
Develop understanding of Explainability, Interpretability, and Trustworthiness of black-box ML/RL models used for dynamical systems and controls.
Emerging topics in Learning-based controls.
Enhance programming and mathematical analysis skills.

Syllabus:

The class syllabus can be found here Download syllabus (PDF)

Products:

Projects:

Deep Transformer Q-Network for Energy Management Strategy for Series Hybrid Agricultural Tractor
DQN-Driven Fusion of Heterogeneous Experts for Intrusion Detection System
Hierarchical Deep Reinforcement Learning-Based Maneuver Decision-Making for Autonomous Driving
Synthetic Data generation for cross-task generalization in Vision-Language-action models
Vision action language Foundation Models for General-Purpose Robotics
A Hybrid Reinforcement Learning and CLF-CBF-QP Framework for Safe Autonomous Vehicle Navigation
Learning Structured Skills for Bipedal Navigation via Mixture-of-Experts Decision Transformer
Learning-Based Powertrain Control for Electrified Powertrains with Integrated Aging of Battery, and Aftertreatment System
Traction Electric Machine Speed Synchronization Under Uneven Torque Allocation for Heavy Duty Electric Vehicle
Deep Koopman Operator Autoencoder Learning for 5-link-walker RABBIT
Quantum Machine Learning for Transportation System Optimization
End-to-end Learning Based Autonomous Vehicle Control via Reinforcement Learning
Safe controller for autonomous driving
Stuart–Landau Actor (SLA): Dynamical Weight Updates for Efficient Exploration in Control Tasks