The size of miniature urban search and rescue robots forces constraints on critical resources such as power, space, and computational ability. To alleviate the effect of resource constraints, it is prudent to let robots adapt themselves autonomously in reaction to unforeseen conditions in disaster area. Software adaptation has been the goal of many researchers, but hardware adaptation is an important aspect, too. In this paper, we present a self-adaptation framework for heterogeneous multi-robot collectives, which consists of partial dynamic reconfiguration for hardware adaptation, module migration for software adaptation, and task performance evaluation to enable autonomous adaptation. These things taken together represent an Embedded Virtual Machine for hardware/software task migration. Our Morphing Crossbar structure, which interconnects reusable hardware modules, is presented to increase the partial dynamic reconfiguration performance. Combined with the runtime downloading and installation of reusable software modules based on the port-based object framework in our PBO/RT realtime operating system, high-performance, heterogeneous, runtime adaptation of hardware/software real-time computational systems is achieved. The adaptation evaluators provide the ability of robotic systems to self-diagnose and self-adapt with an appropriate, distributed configuration monitor. We have implemented this self-adaptation framework in our RecoNode CPU node. By using RecoNode-based CRAWLER and HexRotor robots, an application scenario is presented in order to show the feasibility of this framework.