A reactive navigation method based on an incremental learning of tasks sequences

Frédéric Davesne (1) , Claude Barret (1)

1 CEMIF - Centre d'Etude Mécanique d'Ile-de-France

Within the contest of learning sequences of basic tasks to build a complex behavior, a method is proposed to coordinate a hierarchical set of tasks. Each one possesses a set of sub-tasks lower in the hierarchy, which must be coordinated to respect a binary perceptive constraint. For each task, the coordination is achieved by a reinforcement learning inspired algorithm based on the heuristic which does not need internal parameters. A validation of the method is given, using a simulated Khepera robot. A goal-seeking behavior is divided into three tasks: go to the goal, follow a wall on the left and on the right. The last two tasks utilize basic behaviors and two other sub-tasks: avoid obstacles on the left and on the right. All the tasks may use a set of 5 basic behaviors. The global goal-seeking behavior and the wall-following and the obstacle avoidance tasks are learned during a step by step learning process.

Mots clés: task sequence learning; constraint based learning; perception/action loop

Domaines: Apprentissage [cs.LG]; Robotique [cs.RO]

Frédéric Davesne, Claude Barret. A reactive navigation method based on an incremental learning of tasks sequences. First IEEE Workshop on Robot Motion and Control (RoMoCo 1999), Jun 1999, Kiekrz, Poland. pp.29--34, ⟨10.1109/ROMOCO.1999.791048⟩. ⟨hal-00377772⟩