<p dir="ltr">Robotic software design and implementation have traditionally relied on human engineers to fine-tune parameters, optimize hardware utilization, and mitigate unprecedented situations. As we face more demanding and complex applications, such as distributed robotic fleets and autonomous driving, explicit fine-tuning of autonomous systems yields diminishing returns. To make autonomous systems smarter, a design-time and run-time framework is required to extract constraints from high-level human decisions, and self-adapt on-the-fly to maintain desired specifications. Specifically, for controllers that govern cyber-physical interactions, making them self-adaptive involves two challenges. Firstly, controller design methods have historically neglected computing hardware constraints that realize real-time execution. Hence, intensive manual tuning is required to materialize a controller prototype with balanced control performance and computing resource consumption. Secondly, precisely modeling the physical system dynamics at edge cases is difficult and costly. However, with modeling discrepancies, controllers fine-tuned at design time may fail at run time, causing safety concerns. While humans are inherently adept at reacting and getting used to unknown system dynamics, how to transfer this knowledge to robots is still unresolved.</p><p dir="ltr">To address the two challenges, we propose a designer-augmenting framework for self-adaptive control systems. Our framework includes a resource/performance co-design tool and a model-free controller self-adaptation method for real-time control systems. Our resource/performance co-design tool automatically exploits the Pareto front of controllers, between real-time computing resource utilization and achievable control performance. The co-design tool simplifies the iterative partitioning and verification of controller performance and distributed resource budget, enabling human engineers to directly interface with high-level design decisions between quality and cost. Our controller self-adaptation method extracts objectives and tolerances from human demonstrations and applies them to real-time controller switching, allowing human experts to design fault mitigation behaviors directly through coaching. The objective extraction and real-time adaptation do not rely on prior knowledge of the plant, making them inherently robust against mismatch between the design reference model and the physical system.</p><p dir="ltr">Only with the prerequisite of real-time schedulability under Worst-Case Execution Time (WCET), will the digital controller deliver the designed dynamics. To determine the real-time schedulability of controllers during the design-time iteration and run-time self-adaptation, we propose a novel estimate of WCET based on the Mixed Weibull distribution of profiling statistics and a linear composition model. Our hybrid approach applies to design-time estimation of arbitrary-scaled controllers, yielding results as accurate as a state-of-the-art method while being more robust under small profiling sample sizes. Finally, we propose a resource consolidator that accounts for real-time schedulable bounds to utilize available computing resources while preventing deadline misses efficiently. Our consolidator, formulated as a vector packing problem, exploits different parallelization techniques on a CPU/FPGA hybrid architecture to obtain the most compact allocation plan for a given controller complexity and throughput. </p><p dir="ltr">By jointly considering all four aspects, our framework automates the co-optimization of controller performance and computing hardware requirements throughout the life cycle of a control system. As a result, the engineering time required to design and deploy a controller is significantly reduced, while the adaptivity of human engineers is extended to fault mitigation at run-time.</p>