The CO-MAN project has unified model-based control with machine learning to establish a comprehensive theoretical framework for safe, user-adaptive control. Research achievements include the derivation of Bayesian prediction error bounds for Gaussian Process (GP) regression, which establish a direct link between local training data density and prediction accuracy. These results facilitated the proof that tracking errors asymptotically diminish as the data density increases. The project developed approximated Gaussian Process Dynamical Models (GPDMs) that restore the Markov property while reducing computational complexity. Addressing the challenge of safe learning, the work delivered an episodic algorithm that iteratively generates training data in task-relevant regions to ensure a contraction of the maximum tracking error. Furthermore, a safe event-triggered learning framework was established, allowing for the identification of critical model uncertainty and providing a safe excitation signal to maintain safety filter feasibility without a pre-defined backup controller. Individual user preferences were integrated into control design through human-interactive methodologies, including a control-oriented Inverse Reinforcement Learning (IRL) framework that learns human behaviour from traffic data while maintaining practical stability. In the personalised healthcare domain, a robust calibration framework for Functional Electrical Stimulation (FES) was established using a Student-t likelihood to manage the heteroskedastic and heavy-tailed noise inherent in neuromuscular responses. These theoretical advancements were validated across high-impact scenarios, including the systematic benchmarking of deep learning models for Parkinson’s disease symptom estimation and the validation of user-centric designs in assistive robotics. Further results include physically consistent learning for Lagrangian systems, online detection of compensatory movement strategies in clinical populations, and a computation-aware learning framework to ensure stable control under real-time constraints.
Exploitation of the project results focuses on providing a technically viable path for personalising human-centric control systems through the release of several specialised software repositories. These include accompanying code for risk-sensitive safety filters for reinforcement learning, the PGopt repository for Particle Gibbs-based optimal control, and the koopcore package for Koopman kernel learning. Additional exploitation materials include a framework for stable inverse reinforcement learning based on control Lyapunov landscapes, as well as an implementation for learning geometrically-informed Lyapunov functions with deep diffeomorphic networks. The project also delivered an implementation of Lagrangian-Gaussian Processes for projector-based control and a tool for calculating uniform error bounds for Gaussian Processes with unknown hyperparameters. Dissemination was achieved through participation in or organisation of 28 international scientific conferences and workshops, including prominent venues such as the IFAC World Congress and NeurIPS.
