Behavior recognition for humanoid robots using long short-term memory

Learning from demonstration plays an important role in enabling robot to acquire new behaviors from human teachers. Within learning from demonstration, robots learn new tasks by recognizing a set of preprogrammed behaviors or skills as building blocks for new, potentially more complex tasks. One important aspect in this approach is the recognition of the set of behaviors that comprises the entire task. The ability to recognize a complex task as a sequence of simple behaviors enables the robot to generalize better on more complex tasks. In this article, we propose that primitive behaviors can be taught to a robot via learning from demonstration. In our experiment, we teach the robot new behaviors by demonstrating the behaviors to the robot several times. Following that, a long short-term memory recurrent neural network is trained to recognize the behaviors. In this study, we managed to teach at least six behaviors on a NAO humanoid robot and trained a long short-term memory recurrent neural network to recognize the behaviors using the supervised learning scheme. Our result shows that long short-term memory can recognize all the taught behaviors effectively, and it is able to generalize to recognize similar types of behaviors that have not been demonstrated on the robot before. We also show that the long short-term memory is advantageous compared to other neural network frameworks in recognizing the behaviors in the presence of noise in the behaviors.