Theta Neuron Networks: Robustness to Noise in Embedded Applications

In this paper, we train a one-layer Theta Neuron Network (TNN) to perform a Braitenberg obstacle avoidance algorithm on a Khepera robot. The Theta neuron model is more biologically plausible than the leaky integrate and fire model typically used in Spiking Neural Networks. Our motivation is to determine if the dynamical properties of the theta neuron model can be leveraged to increase the noise robustness in an embedded application. We compare Khepera obstacle avoidance results with traditional Artificial Neural Network and TNN implementations under different levels of sensor noise. As the noise increases, the performance of the TNN is the least affected. At high noise levels, the ANN and Braitenberg implementations calculate the incorrect turn direction 42% more often than the TNN and deviate from a straight path trajectory over 10 times as far. The results demonstrate that TNNs warrants further development for engineering applications.