An Adaptive Neuro-Endocrine System for Robotic Systems Timmis, J., Neal, M. and Thorniley, J. (2009) “An Adaptive Neuro-Endocrine System for Robotic Systems”. In: Workshop on Robotic Intelligence in Informationally Structured Space, RIISS '09. IEEE. pp. 129-136 This paper is apparently the first to present the idea of long term autonomy using a neuro-endocrine homeostatic controller. They’ve developed online learning for this kind of controller and have implemented it on a robot. Apparently the neuro-endocrine architecture is first proposed in this paper: M. Neal and J. Timmis, “Timidity: A Useful Mechanism for Robot Control?” Informatica, vol. 27, no. 4, pp. 197–204, 2003. In the above paper, a neuro-endocrine cell is modelled as a normal artificial neurone with a gland cell attached. The gland cell regulates the amount of hormone present, based sensory inputs and hormone decay. The effect is that the weights into the neurone are each multiplied by some variable which is a factor of the hormone concentration, the sensitivity to that hormone and a “match term” whatever that is. Anyway, this is how neuro-endocrine controllers are “traditionally” made. In this paper what they do is they’re going to build on the work they did in the timidity paper and make a robot with 16 proximity sensors avoid obstacles. They did this in the timidity paper, but this time there is online learning. Section 3: Architecture Design and Development There are 2 signals, the first is a “collision” signal which is active when the robot runs into something, it always activates object avoidance. The second is a “proximity” signal “and initially is activated only by small, random amounts when the proximity sensors provide input data”. What I suppose that means is that the threshold for activating a proximity signal is randomly initialised and as a small amount, I think by small amount they mean that the signal is triggered when a thing is quite far away but is still detectable. As the robot explores, it should learn how to associate the collision and proximity signals, since the proximity signal will always precede a collision. The first step was to make a gland. This is the bit that deals with the concentration levels of the hormone, and is described by the following equation: ) ( ) 1 ( ) ( t R t c t c g g g g     Where c g (t) is the concentration of hormone g at time step t, β g is the decay rate of hormone g and R g (t) is the new hormone put in at time t based on sensory stimulation and is discussed later. The neuron outputs the weighted sum of its inputs, there’s some bias to them and a threshold function is applied.