Classical sensor-based approaches tend to constrain all the degrees of freedom of a robot during the execution of a task. In this paper, a new solution is proposed. The key idea is to divide the global full-constraining task into several subtasks, which can be applied or inactivated to take into account potential constraints of the environment. Far from any constraint, the robot moves according to the full task. When it comes closer to a configuration to avoid, a higher level controller removes one or several subtasks, and activates them again when the constraint is avoided. The last controller ensures the convergence at the global level by introducing some look-ahead capabilities when a local minimum is reached. The robot accomplishes the global task by automatically sequencing sensor-based tasks, obstacle avoidance, and short deliberative phases. In this paper, a complete solution to implement this idea is proposed, along with several experiments that prove the validity of this approach.