A novel method for combination of supervised learning and fuzzy reinforcement learning (FRL) is proposed.

Supervised learning is used for initialization of value (worth) of each candidate action of fuzzy rules in critic-only based FRL algorithms.

The subsumption architecture is used for robot navigation.

The proposed algorithm, called SFSL, is used to drive a real robot (E-puck) in an environment with obstacles.

SFSL outperforms FSL in terms of speed learning, and the number of failures.