CO-MAN: Safe data-driven control for human-centric systems
The CO‑MAN research project aims to develop a novel framework for user‑adaptive, data‑driven control with formal performance guarantees, addressing the scientific challenges posed by high uncertainty and diverse individual user requirements.
Motivation
The EU‑funded CO‑MAN project is developing a new generation of safe, user‑adaptive control systems for applications in which humans are an integral part of the control loop. These human‑centric systems are of increasing importance in domains such as healthcare, mobility, and infrastructure, where technology must operate in close partnership with people to enhance safety, effectiveness, and personalisation. In healthcare, for example, making services more human requires digital systems that enable richer and more meaningful interaction between patients, clinicians, and technology.
CO‑MAN aims to deliver a framework for data‑driven control with formal performance guarantees, capable of adapting to each user’s specific needs and capabilities. The principal challenge is that human behaviour is complex, variable, and difficult to capture using conventional model‑based techniques, yet understanding this behaviour is essential for both safety and efficacy. The project addresses this challenge by combining probabilistic non‑parametric modelling techniques from statistical learning theory with novel risk‑aware control methodologies and incorporating active user modelling to enable real‑time adaptation to the individual.
A central element of the approach is the application of reliable machine learning methods, such as Gaussian Processes and related Bayesian techniques, which explicitly quantify uncertainty in both human behaviour and system dynamics. This capability supports confidence‑driven control strategies that can adapt online while avoiding unsafe actions. The framework will be validated in high‑impact healthcare scenarios, including personalised disease management for Parkinson’s disease and adaptive assistive robotics for rehabilitation.
The transformative potential of CO‑MAN lies in the convergence of recent advances in trustworthy machine learning with theoretical guarantees on learning behaviour. This integration paves the way for super‑individualised, safe, and adaptive control in real‑world human‑in‑the‑loop systems.
Research focus
We conduct research that bridges fundamental control theory and application‑driven design for human‑centric systems. Our work advances safe, adaptive, and uncertainty‑aware control by combining rigorous model‑based methods with data‑driven learning, including probabilistic and physics‑informed modelling. We study the dynamics of human‑in‑the‑loop interaction, addressing challenges such as redundant actuation and stochastic variability in human neuromechanics. These methods are applied in clinical and real‑world contexts to create personalised, closed‑loop systems for rehabilitation, healthcare monitoring, and human‑performance assessment, enabling safe and trustworthy assistance that adapts to individual needs.
Safe learning‑based control and model predictive control
Safe learning‑based control and model predictive control focuses on developing methods that combine the adaptability of data‑driven learning with the formal guarantees of model‑based control theory. The objective is to ensure stability, safety, and performance even when the system dynamics are uncertain or partially unknown. This includes designing risk‑aware and constraint‑satisfying controllers, such as stochastic model predictive control and safe reinforcement learning, that explicitly account for uncertainty in both models and measurements. By embedding rigorous theoretical bounds into learning‑based controllers, these methods enable trustworthy decision‑making in safety‑critical settings.
Probabilistic Modelling and Gaussian Process‑Based Learning for Control
Probabilistic modelling and Gaussian Process‑based learning for control develops uncertainty‑aware, non‑parametric models of dynamical systems that can be seamlessly integrated into control frameworks. Gaussian Processes and related Bayesian methods are used to capture complex, nonlinear system behaviour while providing explicit measures of predictive uncertainty. Where possible, models incorporate physical structure, such as Lagrangian formulations derived from mechanical principles, or mathematical structure, such as Koopman embeddings that offer linear representations of nonlinear dynamics. These structured approaches improve interpretability, sample efficiency, and generalisation. The resulting models form the foundation for robust and adaptive controllers capable of operating reliably under modelling errors, noise, and limited data
Human‑in‑the‑Loop and Assistive Robotics
This research area addresses the fundamental control challenges of systems in which humans and machines share physical interaction. Such systems often exhibit actuation redundancy, for example when voluntary human effort, robotic actuation, and neuromuscular stimulation can all contribute to the same movement. Control strategies must therefore resolve how these sources of actuation are coordinated in real time to maintain stability, optimise task performance, and preserve user engagement. Human neuromechanical behaviour is inherently variable and can be modelled as a stochastic process, with uncertainty arising from sensor noise, physiological variability, and task‑dependent adaptation. We develop adaptive, uncertainty‑aware control methods that integrate multimodal sensing such as EMG, detect and mitigate compensatory movement patterns, and employ physics‑informed neuromechanical models in combination with probabilistic learning. These methods provide robust, personalised control solutions that generalise across a wide range of assistive technologies.
Data‑Driven Control for Applied Domains
This research area focuses on the translation of probabilistic modelling, uncertainty‑aware control, and human‑in‑the‑loop strategies into real‑world applications, with a strong emphasis on clinical rehabilitation and related health‑care technologies. The goal is to design control‑oriented, patient‑ or user‑specific systems that adapt assistance and rehabilitation strategies to individual needs. These systems integrate advanced modelling with real‑time control to ensure that decision‑support is coupled with safe and effective actuation. Applications include estimating Parkinson’s disease symptoms from wearable sensors to adapt therapy parameters, quantifying motor function for personalised rehabilitation planning, modelling recovery progression to adjust assistance levels over time, and monitoring physical performance in sports or workplace safety contexts. By combining the methodological advances from safe learning‑based control, probabilistic modelling, and human‑in‑the‑loop control, this research enables closed‑loop, data‑driven systems that support timely intervention, deliver safe and personalised assistance, and optimise outcomes across a range of applied domains.
EXCELLENT SCIENCE - European Research Council (ERC)
Project title: CO-MAN: Safe data-driven control for human-centric systems
Project number: 864686
Call (part) identifier: H2020-ICT-2019-2
Project period: September 2020 - February 2026
Student members
Student assistant / visiting student
- Arnabesh Das
- Xinyi Shao
- Yassine Kallel
