A Bidirectional Value Alignment Method Based on Behavior Trees

In this paper, we propose a bidirectional value alignment method based on behavior trees, aiming at proposing a more reliable and trusting behavior tree for task completion. In order to create an intermediate between human and robots for value alignment, the value model is established to evaluate behavior trees. The value alignment method is consist of active and passive value alignment, with feedback from human and explanation from robots, combining human intelligence with machine intelligence, instead of complying with human's instructions only. The effectiveness of our method is validated by a treasure hunting game, showing that a reliable and trusting behavior tree can be obtained through the bidirectional value alignment process.