In the recent years, there has been a growing interest in self-adaptive embedded systems. Compared to the conventional approach, they require a control loop based on a three-step process: (1) observation, handled by a set of sensors/monitors, (2) diagnosis, which analyzes observed data to adapt the system, and (3) action, which tunes system parameters accordingly. Putting an additional intelligence into the circuit so that it is capable of modifying itself a set of parameters is not a new idea. But today, it seems that the conditions have been met to build such circuits. Firstly, self-observation has been made feasible with different kind of monitors, like activity counters, temperature sensors, critical path-monitors, etc. Secondly, it is possible to tune the voltage/frequency pairs, to migrate the code of a given task from one processing element to another, to adapt the routing of data in the interconnection network, etc. So what is the real challenge today? Achieving a complex but realistic unified self-adaptation mechanism, which strikes the balance between the introduced overhead, power consumption, performance and area. Given the increasing complexity of embedded systems, our approach is to consider a regular distributed architecture, with a set of identical Processing Elements, interconnected with a network on chip. Thus, all the hardware/software building blocks required for self-adaptation, are the same for each PE, which simplifies the scalability for future technologies. During this talk, we will present an open experimental platform and original approaches for the control loop based on the three-step adaptation process; we will analyze the cost of their implementation and will draw the perspectives offered by such techniques.