Long Short-Term Memory in Recognizing Behavior Sequences on Humanoid Robot

In order for robots to learn more complex behaviors, recognizing primitive behaviors plays a fundamental role. Research has shown that the recognition of primitive behaviors such as basic gestures enables robots to learn more complex behaviors as combinations of these simple, primitive behaviors. The focus of this study is to investigate the tolerance of neural network models to noisy inputs. We compare and evaluate several neural network architectures including the multilayer perceptron (MLP), time-delay neural network (TDNN), recurrent neural network (RNN) and the Long Short-Term Memory (LSTM). We show that the LSTM is superior to other models in terms of its robustness noisy inputs subjected to Gaussian noise.