Abstract
The main argument of this work is that present-day autonomous systems and computing tools too frequently sacrifice human legibility and alignment in favor of technical efficiency. The rapid rise of self-governing systems and advanced computing has created a significant “interpretability gap” between system performance and human needs. In computing, expensive GPU systems are often underused because only a limited number of people can use them efficiently. Traditional scheduling tools are built for experts, so beginners and non-technical users have difficulty with complex interfaces and unclear policies. In assistive robotics, engineering metrics focus on speed, completion, and autonomy, while geriatric care standards value emotional connection, dignity, and shared agency. This divide puts assistive robotics in a position to provide a checklist of care rather than genuine care. The core problem connecting both projects is the risk of “black box” systems that boost technical efficiency yet lessen human understanding. For example, whether a student waits to see when a machine learning job will run, or how it’s running, or an elderly patient tries to keep control over a robotic caregiver, a lack of transparency erodes trust and limits effective use. A student might not understand how their machine learning job ended up in the system’s queue, or an elderly patient might not understand why a robotic caregiver is making certain decisions, or even stop it if it’s wrong. Addressing this shared challenge, these projects emphasize aligning technical capabilities with user requirements, thereby increasing interpretability, understanding, and accessibility. Specifically, designing accessible scheduling interfaces for different tasks and examining discrepancies and divergences in standards for robotic care. The rationale for this project stems from the high cost and poor utilization of GPUs in small-lab or academic environments, where elaborate scheduler interfaces such as Slurm or Kubernetes frequently discourage inexperienced users. These limitations directly reflect the larger issue of accessibility and user trust discussed above. Research shows that static scheduling baselines in these environments often achieve utilization rates of only 45-67%. To address this, the project involves developing a “pop-up” GPU scheduler. It is a temporary, isolated, and containerized service designed for easy deployment on single-host machines without permanent infrastructure. The methods include a web-based UI for intuitive job submission and status tracking, where jobs are submitted from other devices to the host device, using Docker for job isolation to protect the host operating system, and implementing hardware-aware temperature and utilization to pause queues if temperatures exceed safety limits. Evaluation is performed through A/B testing (comparing the UI to traditional CLI tools), system performance logs, and user surveys to assess perceived equity and transparency. The results aim to demonstrate a functional prototype that allows multiple users to successfully queue and execute jobs with minimal configuration. Ultimately, the project concludes that reducing the technical burden through visualization and automated containerization increases the usability of resources and system dependability for non-expert researchers. The next project examines how engineering goals and evaluation metrics for assistive robots compare to geriatric care standards. While engineers prioritize efficiency—speed, accuracy, and autonomy—care ethics emphasize attentiveness and patient agency. To illustrate this difference, the project compares technical benchmarks, such as time-per-task, with the geriatric '5Ms': Multicomplexity, Mind, Mobility, Medications, and What Matters Most. Full autonomy runs the risk of objectifying users and reducing care to a checklist, rather than fostering a relationship. Technical measures alone are insufficient for domestic care. Success requires a shared-autonomy model that respects emotional needs and individual choice. Together, these projects demonstrate that successful automation cannot be measured only by throughput, speed, or technical correctness. Just as a GPU scheduler that obscures queue logic risks becoming efficient but mistrusted, a care robot that ignores patient agency faces similar consequences. Thus, my work suggests that better design requires transparency, consistent feedback, and shared control between humans and machines. By aligning evaluation models with human-centered priorities, developers can create systems that are not merely technically effective but also supportive of the individuals they are designed to assist. Subsequent research should continue to explore how to integrate qualitative measures of care and usability into technical evaluation processes, making sure that as systems become more autonomous, they remain fundamentally accountable to the people who use them. This twofold approach to computing and robotics emphasizes that the value of a system is defined not just by what it can do, but by how well it serves the needs and preserves the dignity of its human collaborators.
