Although robot may seem synonymous with autonomy, robots in the real world rarely work without human interaction. Whether because they need to be re-configured to suit a specific task or user need, or because they are intended to work closely with their users, robot development at all stages---from broad concept to fine-tuned behavior---must learn from people in order to satisfy user needs and expectations. These expectations and concerns, which influence human-robot interactions, are themselves influenced by society's narratives dating back to the origins of robotics.

In this thesis, I discuss two topics in human-in-the-loop robotics. In Part I, I show work on planning and safety for guide robots, a form of assistive technology where blind and low-vision human users interact physically with a robot during navigation. In this area, gathering user requirements is critical but difficult, and personalization is necessary. In Part II, I look at online parameter optimization from human preference information, for which I explore information-geometric methods of black-box optimization. I demonstrate the efficacy of the popular CMA-ES algorithm for fast convergence from noisy signals, and discuss a multimodal algorithm based on Gaussian mixture models. Finally, in Part III, I look at how broader ideas and expectations of robots have been influenced by society as a whole by showing how the origin of the term has roots in cultural narratives about disability.