Problem

Heavy-duty mobile machines (HDMMs) are undergoing two significant transitions: a shift towards electrification to meet climate goals, and a move towards increased autonomy using advanced AI. These transitions present substantial technological, economic, and safety challenges. Current diesel-hydraulic systems are complex to replace, and AI adoption is limited by stringent safety standards and reliability concerns in hazardous operational environments. Specific control design difficulties include managing the high-order, nonlinear, and time-varying dynamics of complex actuation mechanisms like electromechanical linear actuators (EMLAs), handling the tightly coupled interactions in multi-degree-of-freedom robotic manipulators, and dealing with intense uncertainties and external disturbances such as motor torque disturbances and voltage perturbations. A critical issue is the inherent trade-off between system robustness and responsiveness. Furthermore, the lack of interpretability and formal guarantees for black-box deep learning models makes their integration into safety-critical HDMMs deeply problematic, preventing alignment with international safety standards.

Method

This thesis develops a novel, generic control framework centered around a Robust Adaptive Control (RAC) strategy, designed to be independent of the energy source (hydraulic, electric, hybrid). The RAC avoids the "explosion of complexity" of traditional backstepping by treating derivatives of virtual control inputs as uncertainties and guarantees uniformly exponential stability. An observer framework is integrated to accurately estimate system states, including those from noisy sensors or unavailable measurements like velocity, even in the presence of disturbances and unmodeled dynamics. The framework evolves from model-free to model-based RAC by incorporating available subsystem-level modeling information, thereby enhancing responsiveness. To ensure safety, logarithmic Barrier Lyapunov Functions (BLFs) are used to define safety-critical input-output constraints and enforce Prescribed Performance Control (PPC), including dynamic bounds for metrics like overshoot and steady-state error. An intelligent supervisory safety component enhances system availability by managing the robustness-responsiveness trade-off. For integrating Deep Neural Network (DNN) policies, a two-layer hierarchical control architecture is proposed: a low-level safety layer switches from the DNN to the RAC during severe disturbances to maintain safe operation, while a high-level BLF-based safety layer initiates a complete system shutdown if disturbances become unmanageable. The JAYA optimization algorithm is employed for automatic tuning of control and observer parameters, simplifying the design process.

Results

The proposed control framework was rigorously validated through extensive experiments on various heavy-duty machinery. For EMLAs, the RAC demonstrated robust control and uniform exponential stability, managing input constraints, uncertainties, and loads up to 76 kN with position tracking errors below 2mm. On heavy-duty hydraulic In-Wheel Drive (IWD) actuators, the model-free RAC achieved superior velocity tracking compared to benchmark model-free methods. In multi-degree-of-freedom robotic manipulators, the RAC effectively executed tasks despite joint faults, exhibiting uniformly exponential stability and improved tracking when tuned with the JAYA algorithm. For fully electrified n-DoF EMLA-driven manipulators, the model-based RAC, incorporating adaptive networks for disturbance compensation and state observation, outperformed state-of-the-art controllers in tracking accuracy, torque efficiency, and convergence under varying loads and velocities. On a 6000 kg hydraulic wheeled HDMM, a torque-observer-based model-based RAC with BLF-based safety layers achieved superior performance and safety-aware control on challenging terrains (snow, ice), where a standard PID controller failed. Finally, a hierarchical control framework synthesizing DNNs with RAC on a hybrid skid-steer HDMM demonstrated uniformly exponential stability and improved system availability under both nominal and perturbed conditions.

Implications

This work significantly advances the state-of-the-art in nonlinear control and robotics by introducing a unified Robust Adaptive Control (RAC) framework with guaranteed performance. It lays a crucial foundation for accelerating the transition of heavy-duty mobile machinery (HDMMs) towards electrification and higher autonomy, ensuring both safety and reliability. The developed control architecture, combining RAC with safety-guaranteed supervisory elements and learning-based policies, provides a generic and adaptable solution applicable across diverse actuation mechanisms and system configurations, ensuring long-term resilience and compatibility with future technologies. The consistent experimental validation demonstrates superior tracking accuracy, faster convergence, and reduced control effort compared to existing methods, even under intense disturbances and fault conditions. The integration of safety layers ensures robust and stable operation, even with uninterpretable learning-based policies, aligning with strict international safety standards. Future research directions include unifying manipulator and wheeled platform control strategies, extending the framework to higher-order system models, including the full power delivery path, expanding learning-based technologies for HDMMs, further improving system reliability, availability, maintainability, and safety (RAMS), and integrating kinematic-level modeling, perception, and localization for achieving full autonomy.