Synthesis of Deep Neural Networks with Safe Robust Adaptive Control for Reliable Operation of Wheeled Mobile Robots

Abstract

Deep neural networks (DNNs) can enable precise control while maintaining low computational costs by circumventing the need for dynamic modeling. However, the deployment of such black-box approaches remains challenging for heavy-duty wheeled mobile robots (WMRs), which are subject to strict international standards and prone to faults and disturbances. We designed a hierarchical control policy for heavy-duty WMRs, monitored by two safety layers with differing levels of authority. To this end, a DNN policy was trained and deployed as the primary control strategy, providing high-precision performance under nominal operating conditions. When external disturbances arise and reach a level of intensity such that the system performance falls below a predefined threshold, a low-level safety layer intervenes by deactivating the primary control policy and activating a model-free robust adaptive control (RAC) policy. This transition enables the system to continue operating while ensuring stability by effectively managing the inherent trade-off between system robustness and responsiveness. Regardless of the control policy in use, a high-level safety layer continuously monitors system performance during operation. It initiates a shutdown only when disturbances become sufficiently severe such that compensation is no longer viable and continued operation would jeopardize the system or its environment. The proposed synthesis of DNN and RAC policy guarantees uniform exponential stability of the entire WMR system while adhering to safety standards to some extent. The effectiveness of the proposed approach was further validated through real-time experiments using a 6,000 kg WMR.