A smart engineering system requires that its all components can interact with their environment and adapt to changes in both space and time. From this point of view, assistive device systems (ADSs) differ from the usual robotics systems in that the interaction between users and devices in ADSs is generally much more direct and closer, so that users and devices should be able to mutually adapt to each other. This inspired our approach, which takes advantage of not only devices′ adapting to users but also users′ adapting to devices by trial and error, based on the device–user interaction. Several problems of how to realize such mutually adaptive assistive device systems would be discussed, and some of our ideas and approaches would be introduced. The efficiency of the interaction based learning method was verified by its application to two assistive device control tasks: a myoelectrical hand control for wrist amputees, and an Electromyogram (EMG) automatic Functional Electrical Stimulation (FES) switching for gait restoration for hemiplegics. Results were shown with discussion.