Mobile robots and many edge AI devices have a need to trade off computational power against power consumption, battery size, and time between charges. Consequently, it is common for such devices to have significantly less computational power than the powerful GPU-based systems typically used to train and evaluate deep neural networks. Object detection is a key aspect of visual perception for robots and edge devices but popular object detection architectures that run fastest on GPU based systems or that are designed to maximize mAP with large input image sizes may not scale well to edge devices. In this work we evaluate the latency and mAP of several model architectures from the YOLO and SSD families on a range of devices representative of robot and edge device capabilities. We also evaluate the effect of runtime framework and show that some unexpected large differences can be found. Based on our evaluations we propose new variations of the YOLO-LITE architecture which we show can provide increased mAP at reduced latency.