Lower-limb exoskeletons have potential to increase endurance, reduce incidence of injury, aid rehabilitation and prolong independence. Progress in exoskeleton research is slowed by the complex nature of exoskeleton design and by the complexity of the humans they are supposed to assist. Developing highly versatile tools to address these issues will accelerate research and bring exoskeletons closer to becoming life changing products. This thesis details an approach to exoskeleton development that leverages both universal exoskeleton emulators and human-in-the-loop optimization to enable accurate comparisons of qualitatively different exoskeleton assistance strategies. Instead of building new or adjusting old hardware to test a given exoskeleton design, we developed highly capable, lightweight, tethered exoskeletons that were used to emulate the behavior of a wide variety of possible exoskeleton designs. The development of these exoskeleton emulators revealed guiding principles of lower-limb exoskeleton design that will aid in the creation of future exoskeletons. Experimental comparisons of exoskeleton control architectures for walking and running were performed to determine the most promising strategies to pursue. Passive, spring-like assistance was compared to powered assistance for running. Three qualitatively different assistance strategies were explored for assisting walking. Included among the walking assistance strategies was a first look at an EMG-based muscle-tendon controller. This controller applied a muscle-tendon model to measured EMG activity to produce exoskeleton assistance in a manner similar to human muscle. In both the walking and running experiments, human-in-the-loop optimization was performed to provide customization of exoskeleton settings to fit the needs of individual subjects. This enabled a more accurate comparison of candidate assistance strategies. We expect our findings to influence the design of portable, untethered ankle exoskeletons and to proliferate the use of emulator systems in combination with human-in-the-loop optimization to perform side-by-side comparisons of assistance strategies.